jain.com — data centers, connectivity and security, analyzed daily

Author: Deepak Jain

  • 800V AI Rack Power Is Being Built in Layers: Navitas GaN, Microchip Control

    800V AI Rack Power Is Being Built in Layers: Navitas GaN, Microchip Control

    TL;DR · 30-second read

    The Short Version

    Microchip and Navitas, two chip companies, have published a blueprint for a power board that feeds the chips running artificial intelligence. It takes electricity at 800 volts, several times what a home outlet supplies, and steps it down in a single move to the 6 volts those chips use.

    The board is about a fifth thinner than a phone. At its best, it aims to lose just 4% of the power along the way.

    Artificial intelligence buildings are straining for electricity. This fix is being assembled from more than one company’s parts.

    Microchip Technology and Navitas Semiconductor have jointly released an 800V DC-to-6V DC reference design for AI data center rack power, ET Datacenters reported on October 10, 2026. The design pairs Microchip’s dsPIC33AK digital signal controllers and TA100 security chip with 16 of Navitas’ GaNFast gallium nitride transistors. It targets up to 96% peak efficiency at full load and a power density of 2,100 watts per cubic inch, and the companies say it aligns with the Open Compute Project’s 800V DC standard.

    The release arrives four days after Navitas, in an 8-K filed October 6, reported closing its acquisition of Claros, Inc., a developer of integrated voltage regulators that handle the final step of delivering power to AI processors.

    Executive Summary

    A reference design is an engineering blueprint: working hardware, software and documentation that other companies can copy or adapt instead of starting from scratch. This one tackles a specific problem in AI racks: converting high-voltage 800V DC power down to the low voltage that sits near the GPU board. It does so in a single conversion stage, where conventional designs use two.

    The design matters for two reasons. First, it puts concrete targets on the 800V transition: 96% peak efficiency, 2,100 W/in³ and a board the companies describe as about 20% thinner than a mobile phone. Second, it divides the work among specialist suppliers. Navitas supplies the power switches. Microchip supplies the controller that times them and the chip that secures the firmware. As hyperscalers and power-system makers move toward 800V racks, that layered structure points to a supply chain with more than one chipmaker in each power board.

    The targets are design goals, not field results. The companies have not named customers, production timelines or pricing.

    Three Jobs, Two Companies: How the Board Divides the Work

    Converting 800 volts down to 6 volts in one step requires three distinct kinds of silicon, and this design takes them from two suppliers. The first job is switching. The power delivery board uses 16 Navitas NV6034 GaN transistors, each rated at 650 volts and 17 milliohms of resistance, in a stacked half-bridge arrangement. That arrangement spreads the high input voltage across devices individually rated below it. Gallium nitride is favored here because it can switch on and off very quickly with little loss, and the design runs at 1 MHz: one million switching cycles a second.

    The second job is control. At 1 MHz, each switching cycle lasts one microsecond, and the controller has to decide every cycle how long each transistor stays on. Microchip’s dsPIC33AK256MPS306 family brings a 200 MHz 32-bit core with a double-precision floating-point unit, pulse-width modulators with 78-picosecond resolution, and 12-bit analog-to-digital converters sampling at up to 40 million samples per second. Put plainly, the controller can adjust timing in steps of roughly one twelve-thousandth of a single cycle while measuring voltage and current fast enough to react within it. The third job is trust. Microchip’s TA100 provides a hardware root of trust for secure boot, authenticated firmware updates and key management.

    This division is what makes the transition multi-vendor in practice. GaN expertise, real-time digital control and hardware security are separate engineering disciplines. A reference design that pre-validates one combination of them lowers the cost for power-shelf makers and hyperscaler hardware teams to adopt that combination. The companies say the package reduces design risk and shortens development cycles. Because the design aligns with the Open Compute Project’s 800V DC standard, an open specification that rack buyers use to compare and interchange suppliers, it also sits inside a framework built for multiple vendors rather than one. This is a single data point, not proof of an industry pattern. Still, it shows how the parts list for 800V rack power is taking shape.

    A Grid-to-xPU Strategy That Still Partners at the Control Layer

    The timing is notable. Navitas’ October 6 filing frames its strategy as covering power delivery from the grid to the processor. The company describes its GaN and high-voltage silicon carbide devices as enabling the new 800V DC architecture, and the newly acquired Claros technology as the last step to the chip. Navitas told investors it expects Claros to add at least $3.5 billion to its identified 2030 serviceable available market, more than doubling the total to over $8 billion. That total combines the existing $3.5 billion GaN and silicon carbide opportunity with about $1 billion from new JFET technology.

    Yet in this design, the control and security silicon comes from Microchip. That does not contradict Navitas’ strategy. Claros’ integrated voltage regulators, as the filing describes them, address a different stage of the power path. What the pairing does show is that even a supplier assembling an end-to-end portfolio is working with partners where another company has an established controller and security line. For buyers, that is mildly reassuring: the 800V transition is not funneling every component through one chipmaker. For Navitas, the partnership is a channel. A Microchip-backed reference design can put GaNFast devices in front of engineers who already standardize on Microchip controllers.

    The serviceable-market figures deserve to be read as what they are: Navitas’ estimate of the opportunity its products could address by 2030. They are not revenue, backlog or contracted demand.

    One Stage Instead of Two: The Efficiency Argument

    The core technical claim is that converting 800V DC directly to 6V DC folds two conversion stages into one converter, improving end-to-end efficiency. The logic is straightforward. Every conversion stage loses some energy as heat, so removing a stage removes a source of loss. Distributing power at 800 volts helps for a related reason: higher voltage means less current for the same power, which cuts resistive losses and the amount of copper needed in the rack.

    The 96% figure is a peak target at full load, and the release discloses no efficiency curve. Converters typically perform differently at partial load, where AI racks may spend much of their time, so operators will want measured data across the operating range. The physical design may matter as much as the efficiency. A board about 20% thinner than a phone can sit close to the GPU board, which the companies say improves transient performance. Transient performance is the converter’s ability to respond when processor demand jumps suddenly. Shorter distances between converter and chip mean less voltage drop and faster response, a growing concern as Navitas’ own CEO letter describes processors needing thousands of amps delivered with high precision.

    Security Moves Into the Power Path

    Including a dedicated security chip in a DC/DC converter design reflects a quieter shift. Rack power hardware now carries firmware that controls how electricity reaches some of the most expensive processors in a data center, and that firmware can be updated. The TA100’s secure boot and authenticated updates are meant to ensure only trusted code runs on the converter. The dsPIC33AK family’s support for the post-quantum algorithms recommended in the Commercial National Security Algorithm Suite 2.0 extends that protection against future code-breaking methods.

    For data center security teams, this places power conversion alongside servers and network gear as equipment with a firmware supply chain to manage. That is a practical consequence of the layered design, and it gives procurement teams one more vendor relationship to qualify.

    Background

    Navitas Semiconductor, listed on Nasdaq as NVTS, develops power semiconductors based on gallium nitride and high-voltage and ultra-high-voltage silicon carbide. Under its Navitas 2.0 strategy, the company has refocused on AI infrastructure. On October 6, 2026, it closed its acquisition of Claros, whose integrated voltage regulator technology moves power conversion to within millimeters of the processor. Navitas describes the combined portfolio as covering the full path from grid to processor and has scheduled an October 20 webinar to present it.

    Microchip Technology is a long-established supplier of microcontrollers and digital signal controllers, the chips that run real-time control loops in power supplies, motors and industrial equipment, along with dedicated security devices. The joint design reflects a broader move in AI data centers toward 800V DC rack power, driven by rack power demands that conventional lower-voltage distribution struggles to serve efficiently.

    Sources

    Source: Microchip, Navitas partner on 800V DC power for AI data centers, ET Datacenters: the companies’ 800V DC-to-6V DC reference design for AI rack power.

    Primary sources: Navitas Semiconductor Form 8-K, filed October 6, 2026; Exhibit 99.1: Navitas Closes Acquisition of Claros, Advancing AI Infrastructure with Grid-to-xPU Power Delivery; Exhibit 99.2: A Letter from Chris Allexandre, President and CEO, Navitas 2.0: Breaking the AI Infrastructure Power Wall.

  • A $2B GlobalFoundries-TSMC Deal Onshores AI’s Interposer, Not the Whole Package

    A $2B GlobalFoundries-TSMC Deal Onshores AI’s Interposer, Not the Whole Package

    TL;DR · 30-second read

    The Short Version

    Two major chipmakers, GlobalFoundries and Taiwan Semiconductor Manufacturing Company, the world’s largest contract chipmaker, have struck a $2 billion deal to make a key chip part in the United States.

    The part is a silicon “interposer”: a thin slab of silicon that sits under an artificial intelligence processor and wires it to its memory, like a shared floor both chips are mounted on. The packaging step built around it has been one of the main limits on how many artificial intelligence chips get shipped.

    Making it in America could add supply outside Taiwan. But it covers one step, not the whole chip.

    Yahoo Finance reported that GlobalFoundries and TSMC have reached a $2 billion deal for silicon interposers produced in the United States. Silicon interposers are the base layer in the advanced packages that join high-end AI processors to stacked high-bandwidth memory, which places the agreement inside the part of the GPU supply chain that has repeatedly constrained output.

    The deal follows TSMC’s October 8 Form 6-K, its September 2026 revenue report, which showed consolidated revenue of roughly NT$511.86 billion for the month, up 54.6 percent from September 2025, and NT$3,898.73 billion for January through September, up 41.1 percent year over year.

    Executive Summary

    The headline facts are simple: a $2 billion agreement between the world’s largest contract chipmaker and a US-headquartered specialty foundry, covering a component that is made in the United States. The interposer is not a commodity part. It is a large piece of silicon, patterned in a wafer fab with fine wiring and vertical connections called through-silicon vias. AI accelerators mount their compute die and memory stacks on it side by side.

    That matters because AI chip supply has been gated less by the ability to make processors than by the ability to package them with memory. Making interposers in the US adds a domestic source for one input to that packaging flow. However, an interposer is an intermediate part. Until the companies say where the rest of the package is assembled, the deal shows where part of the bottleneck can be relieved, not that it has been.

    For GlobalFoundries, a $2 billion commitment tied to AI packaging is a significant new line for a foundry that stepped back from the leading-edge process race in 2018. For TSMC, whose September revenue alone was more than NT$500 billion, the deal is small in financial terms. Its importance is geographic and logistical.

    Why the Interposer Sits on AI’s Critical Path

    The most capable AI accelerators are not a single chip. They are assemblies. A large processor die sits next to several stacks of high-bandwidth memory (HBM), DRAM chips stacked vertically to deliver very fast data access. These parts have to talk to each other across thousands of extremely fine connections, far denser than an ordinary circuit board can carry. The silicon interposer solves this. It is a slab of silicon wired with lithography, the same patterning process used to make chips, so the processor and memory can be packed millimetres apart and connected at chip-level density.

    TSMC’s version of this approach is CoWoS, short for chip-on-wafer-on-substrate. Dies are bonded onto an interposer wafer, which is then mounted on a conventional package substrate. In recent years, CoWoS capacity has been widely identified as a binding constraint on AI accelerator shipments. The processors could be fabricated, but there was not enough capacity to package them with their memory. Interposers are one input to that flow. Because they are made in a fab rather than an assembly house, adding interposer supply requires wafer capacity with the right tooling, not just more packaging lines.

    What Onshoring One Layer Does, and What It Doesn’t

    This is the substance behind the headline. A US-based interposer source moves one fab-made step of AI packaging onto American soil. That has real value. It adds a source outside Taiwan for a component tied to a known chokepoint. It suits a mature, non-leading-edge process, which plays to GlobalFoundries’ strengths in its New York and Vermont fabs. It also gives chip designers and their hyperscale customers, the largest cloud and AI operators, a domestic option to cite when they are asked about supply resilience.

    But an interposer alone does not make a finished accelerator. After the interposer is made, the processor dies and memory stacks must be bonded onto it. The assembly must then be mounted on a substrate, tested, and shipped. If those steps remain in Taiwan, US-made interposers simply cross the Pacific in the opposite direction from the processors. Much of the processor output from TSMC’s Arizona fabs has itself needed to return to Taiwan for advanced packaging. The deal becomes a genuinely domestic AI supply chain only when it pairs with US assembly capacity. TSMC said in March 2025 that its expanded Arizona plans include advanced packaging facilities, and its October 6-K shows continued financial support for US subsidiaries, recording lending to TSMC Washington and guarantees for TSMC Arizona. Connecting the interposer supply to that assembly capacity is the step that would turn a single onshored layer into an onshored package.

    The packaging variant also matters. TSMC’s family includes the original CoWoS-S, which uses a full silicon interposer, and newer variants such as CoWoS-L, which replace the full interposer with smaller embedded silicon bridges. If flagship accelerators continue shifting toward bridge-based designs, the market for full-size interposers looks different. Which products these US interposers serve will determine how much of the bottleneck they actually touch.

    Reading $2 Billion Against Each Company’s Scale

    For TSMC, the figure is modest. Its September revenue was about NT$511.86 billion, and its year-to-date revenue was NT$3,898.73 billion, up 41.1 percent. A $2 billion agreement is not a material financial event at that scale. It is a capacity and location decision. Choosing a domestic partner for a mature-process component frees TSMC to focus its own capital elsewhere while adding a US element to its packaging chain.

    For GlobalFoundries, the calculus runs the other way. A foundry that chose in 2018 to stop pursuing the most advanced process nodes now has a direct line into the most supply-constrained segment of the AI hardware market, using processes it already runs. That makes it a significant win, though its annual revenue impact cannot be assessed until the contract’s duration and ramp are known. The two companies were once adversaries in a patent dispute that they settled in 2019 with a broad cross-license. A supply relationship on AI packaging is a notable evolution from that history.

    Background

    TSMC, founded in Taiwan in 1987, is the world’s largest contract chipmaker. It manufactures processors for most leading AI chip designers, and its CoWoS packaging has become central to AI accelerator supply. It is building fabs in Arizona and has said its US plans include advanced packaging. GlobalFoundries, spun out of AMD in 2009, runs fabs in Malta, New York; Essex Junction, Vermont; Dresden; and Singapore. It stopped pursuing the most advanced process nodes in 2018 to focus on specialty technologies.

    The two companies settled a patent dispute in 2019 with a broad cross-license. Their new interposer agreement places GlobalFoundries’ US mature-process capacity inside the AI packaging flow that TSMC dominates.

    Sources

    Source: GlobalFoundries and TSMC reach $2B deal for US-based silicon interposers (Yahoo Finance): report of a $2 billion agreement covering US-made silicon interposers.

    Primary sources: Taiwan Semiconductor Manufacturing Company, Form 6-K filed 2026-10-08 (TSMC September 2026 Revenue Report).

  • NOARK’s 75-Day Pitch Bets AI Data Center Power Gear Sells on Schedule, Not Price

    NOARK’s 75-Day Pitch Bets AI Data Center Power Gear Sells on Schedule, Not Price

    TL;DR · 30-second read

    The Short Version

    • Artificial intelligence is driving a boom in giant computer warehouses, and each one needs heavy electrical equipment before it can switch on.
    • NOARK Electric, a maker of that equipment, says it got one North American order from purchase to job site in about 75 days, roughly two and a half months.
    • Its sales pitch is about speed and meeting local safety standards, not about being cheaper.
    • It has not named the customer or said what that order contained, so it is one example, not proof.

    NOARK Electric said in an Oct. 10, 2026 release distributed via PR Newswire that it showcased its data center power infrastructure lineup at Yotta 2026 in Las Vegas. The equipment spans medium-voltage gas-insulated switchgear, low-voltage switchboards and busway, and air and molded-case circuit breakers, which the company is marketing under a “global-to-local” model that pairs global engineering and manufacturing with North American assembly, compliance work and service.

    The release’s one hard data point: on a recent, unnamed North American project, NOARK says coordinated engineering, manufacturing, logistics and local fulfillment took the order from purchase to on-site delivery in approximately 75 days.

    Executive Summary

    NOARK Electric is positioning itself to win electrical-equipment orders from AI-era data center developers by competing on delivery time and regulatory readiness. The company describes a four-step model, “Engineer. Source. Qualify. Deliver,” in which global R&D and factories supply capacity while a North American team handles application engineering, UL and IEEE compliance, localized assembly, fulfillment and after-sales support.

    The announcement matters because electrical gear sits on the critical path of every data center build: a facility cannot power its servers until switchgear and distribution equipment are installed and commissioned. NOARK’s release frames customer priorities as speed to power, reliability and scalability, and never mentions price, a clear signal that it believes schedule certainty is now the deciding factor in these procurements.

    The supporting evidence is thin but specific: a single 75-day order-to-delivery example, plus product-level claims about where assembly and supply are localized. How far that example generalizes depends on details NOARK has not yet disclosed.

    Price Never Appears in NOARK’s Pitch

    NOARK builds its case around three customer priorities (speed to power, reliable performance and scalable infrastructure), and price is not among them. The only hard number the company offers is a schedule: roughly 75 days from order to on-site delivery on a recent North American project. That choice of evidence is the strategy. A supplier competing on cost would lead with cost. NOARK leads with time.

    The mechanism is straightforward. Switchgear, switchboards and circuit breakers sit on the critical path of a data center build. Until they are installed and commissioned, the building cannot take power from the utility or distribute it to servers and cooling. On an AI campus, where the computing hardware is itself a large capital outlay, every week that hardware waits on electrical equipment is a week of deferred revenue. In that setting a developer can rationally accept a higher equipment price in exchange for a firmer delivery date. NOARK’s own framing, that “power availability, equipment lead times and deployment certainty are increasingly central to project planning,” is a bet that buyers now weigh the trade-off this way.

    The people this targets are developers and EPC (engineering, procurement and construction) contractors who own the schedule, and the engineering firms that write equipment specifications. The limits are just as clear. The 75-day figure describes one project. NOARK has not said what equipment it covered, at what voltage class or quantity, or whether it was a standard catalog item or a custom-engineered lineup. As evidence, it illustrates the pitch. It does not establish a repeatable lead time.

    Compliance Is Where a Challenger Loses Time

    For a supplier that is not already on an engineer’s approved list, the slowest step can be qualification, which means getting products accepted against the codes and standards a project must meet. NOARK addresses this directly. It says its local engineering and compliance team integrates UL (Underwriters Laboratories, the North American safety certification body) and IEEE (Institute of Electrical and Electronics Engineers) requirements into product qualification and project execution, to “reduce qualification uncertainty.”

    That is the second half of the lead-time argument. A fast factory helps little if a product spends months in submittal review or needs field modifications to pass inspection. By promising compliance “from the start,” NOARK is aiming at the hidden schedule risk of switching suppliers, not only at shipping time.

    The specific claim is narrower than the framing. The release says a product platform shown at Yotta was “developed to meet applicable IEEE requirements.” It does not list which products carry UL listings, or for which applications. For specifiers, that product-by-product detail determines whether the compliance advantage holds on a given job.

    A Hybrid Supply Chain Built for Multi-Phase Campuses

    NOARK describes a layered supply model. Low-voltage switchboards and busway (prefabricated enclosed conductors that carry power along data hall rows in place of cable) draw on U.S.-based supply to “shorten supply paths.” Its ACB and MCCB platforms, meaning air circuit breakers and molded-case circuit breakers, are assembled in North America on global designs. The company’s global manufacturing footprint serves as reserve capacity for large, multi-phase deployments.

    That structure fits how AI campuses are built: in phases, with repeated, standardized electrical blocks. Local assembly can shorten the final leg and make change orders easier to absorb, while global plants can supply volume when a later phase arrives. NOARK also extends its local presence to commissioning, technical assistance and after-sales support. This matters because reliability in a live facility depends on service response as much as on hardware.

    The trade-off is dependency. The release says only “selected equipment” is supplied locally, which suggests much of the content still comes through the global network, so delivery certainty still depends on that network performing. The medium-voltage gas-insulated switchgear, the heaviest item in the lineup, is described in terms of compact footprint and global R&D, not local supply. Buyers weighing the model should map which parts of a lineup take the short path and which take the long one.

    Background

    NOARK Electric makes low- and medium-voltage electrical distribution, protection and control equipment for commercial, industrial, infrastructure and data center applications, and operates in North America and other markets. Its data center lineup covers the main electrical path inside a facility: medium-voltage switchgear that receives and controls incoming power, low-voltage switchboards and busway that distribute it, and circuit breakers that protect individual circuits.

    Yotta is a data center industry event held in Las Vegas. NOARK used the 2026 edition to argue that the rapid AI-driven expansion of North American data centers has made power infrastructure, specifically equipment availability, lead times and code compliance, a deciding factor in how quickly projects move from planning to operation.

    Sources

    Source: NOARK Combines Global Scale with Local Execution for Data Center Power Infrastructure, NOARK Electric’s Oct. 10, 2026 release on the data center power equipment and delivery model it showcased at Yotta 2026.

  • $2.5B AI Server Diversion Plea Tests Export Controls Where GPUs Become Servers

    $2.5B AI Server Diversion Plea Tests Export Controls Where GPUs Become Servers

    TL;DR · 30-second read

    The Short Version

    A contractor who worked with Super Micro, an American company that builds the powerful computers used to run artificial intelligence, has admitted in court to helping send those machines to China illegally. The scheme was valued at $2.5 billion.

    The computers held chips from Nvidia, whose chips power much of today’s AI. US rules bar selling those advanced chips to China without government permission.

    Why it matters: the case targets the people between the chipmaker and the final buyer, not the chipmaker itself. Enforcement is reaching further down the supply chain.

    A contractor tied to Super Micro Computer has pleaded guilty in a scheme to divert AI servers containing Nvidia chips to China, Reuters reported. Coverage puts the scheme’s value at $2.5 billion and describes the contractor as a co-defendant of a Super Micro co-founder.

    The plea comes in the same week Super Micro disclosed board changes in an 8-K filed October 8. That filing does not mention the export case.

    Executive Summary

    A guilty plea in a $2.5 billion AI server diversion scheme is among the clearest signals yet of where US export-control enforcement on AI hardware is landing. The equipment at issue was not loose chips but complete servers built around Nvidia GPUs (graphics processing units, the processors that do most AI computation). The person admitting guilt is a contractor, not an employee of the chipmaker.

    That matters because it shifts the compliance burden. Export restrictions on advanced AI chips follow the chip into whatever system it is installed in. That makes everyone who builds, resells or ships those systems a potential point of failure, and of liability. For server makers, distributors and the data center operators who buy from them, the case is a reminder that end-user screening is part of the product, not paperwork around it.

    The plea establishes guilt only for the person who entered it. Anything alleged against other defendants remains an allegation, and no charges against Super Micro as a company have been reported.

    Where a GPU Becomes a Server, Control Gets Harder

    Advanced AI chips leave Nvidia in large, traceable allocations to a relatively small set of server makers and cloud buyers. Once a GPU is installed in a server, it becomes one component inside a finished system. That system can then pass through distributors, contractors and resellers before reaching its final operator. US export rules still apply to the chip inside the box, but the chipmaker’s line of sight usually ends at its direct customer.

    That gap is the mechanism this case exposes. The diverted goods were Super Micro AI servers with Nvidia chips inside, and the guilty party is a contractor in the server maker’s orbit. At a reported $2.5 billion, the scheme shows how much hardware value can move through the integration and resale layer, the stage after the chip is sold but before the system is plugged in. Controls written around the chip are only as strong as the checks at that downstream stage.

    Those affected are the companies that sit between foundry and data hall: server builders, value-added resellers, logistics contractors, and the compliance teams that sign off on who the end user really is. Nvidia’s chips were in the boxes, but Nvidia is not reported to be a defendant and has made no material SEC filing on the matter. The enforcement weight in this case sits downstream.

    What the Channel Now Has to Prove

    For server makers and their partners, export compliance means knowing the end user and the end use, not just the immediate buyer. A diversion scheme works by putting a permissible-looking customer between the seller and a prohibited destination. A plea at the contractor level gives every company in that chain reason to scrutinize intermediaries, shipping patterns and resale terms more closely.

    The economic effect is friction. Legitimate buyers of AI servers, from enterprise IT departments to colocation tenants, may face more documentation about where systems will be deployed and who will operate them. For integrators competing on speed of delivery during a GPU supply crunch, slower deal reviews carry a real cost. The case makes clear that the alternative cost is higher.

    What Is and Isn’t Established About Super Micro

    Super Micro is one of the largest builders of Nvidia-based AI servers, so any case involving its products draws attention to the company. The facts on the record are narrower than the headlines. A contractor has admitted guilt. A co-founder is described as a co-defendant, which is an allegation, not a finding. No charges against the company have been reported.

    Super Micro’s most recent disclosure is an 8-K covering the appointment of former Intel IT chief technology officer Shesha Krishnapura as an independent director and the retirement of director Judy Lin. The filing states that Lin’s departure was not the result of any disagreement with the company, and it does not reference the case. Reading a connection into the timing would go beyond what either the plea or the filing supports.

    Background

    Since 2022 the United States has restricted exports of advanced AI chips and the systems built around them to China, through rules administered by the Commerce Department’s Bureau of Industry and Security and tightened in later rounds. Nvidia, the dominant supplier of AI accelerators, sells its GPUs largely through server makers and large cloud buyers, which turn them into finished systems.

    Super Micro Computer, based in San Jose, California, is one of the biggest of those server makers and a major builder of Nvidia-based AI systems for data centers. Its fiscal year ended June 30, 2026. In early October it announced a board change, adding former Intel IT chief technology officer Shesha Krishnapura and accepting the retirement of director Judy Lin.

    Sources

    Source: Super Micro contractor pleads guilty in scheme to divert AI servers with Nvidia chips to China (Reuters via Investing.com): a contractor pleads guilty in a reported $2.5 billion scheme to divert Nvidia-equipped AI servers to China.

    Primary sources: Super Micro Computer, Form 8-K filed October 8, 2026 (board appointment of Shesha Krishnapura and retirement of Judy Lin).

  • CoreWeave’s 15x Reload Claim Puts AI Inference on the Training Critical Path

    CoreWeave’s 15x Reload Claim Puts AI Inference on the Training Critical Path

    TL;DR · 30-second read

    The Short Version

    CoreWeave is a company that rents out the powerful computer chips used to build artificial intelligence. It has now launched Forge, a software toolkit for building and running artificial intelligence systems. Anyone can start using it for free.

    The surprising part: when these systems learn by trial and error, they keep swapping in updated versions of themselves. CoreWeave says it made that swap 15 times faster in testing, so the learning stops stalling.

    The move also puts CoreWeave in more direct competition with Amazon, Google and Microsoft, which already sell similar toolkits.

    CoreWeave has launched CoreWeave Forge, a platform that connects model serving, observability, post-training and evaluation in one environment. It was unveiled at the Fully Connected event and is free to start, with paid tiers for additional capabilities. CoreWeave’s product page describes Forge as a way to “run, observe, curate, improve, and evaluate models and agents” without locking work to one cloud, model or framework.

    SiliconANGLE reported on October 8 that Urvashi Chowdhary, CoreWeave’s vice president of product and AI services, said a preview capability called CoreWeave RL Rollouts, built on Nvidia’s Dynamo framework, improved model reload latency by 15x compared with a baseline configuration in testing.

    Executive Summary

    CoreWeave built its business renting GPU capacity, the graphics processors that train and run AI models. Forge adds managed software on top of that capacity. According to CoreWeave’s product page, it covers fine-tuning, managed reinforcement learning, per-token serverless inference, dedicated inference, agent tracing, sandboxes, a model and dataset registry, and experiment tracking through Weights & Biases Models. The free entry tier extends the platform to individual developers, not only the large labs and enterprises that sign capacity contracts.

    The most technically significant claim concerns reinforcement learning, the training method used to build agentic models that take multistep actions. In that workflow, the inference system that generates the model’s attempts has to keep reloading freshly trained versions of the model. CoreWeave says RL Rollouts cut that reload time 15-fold in testing. If that holds outside the lab, inference performance becomes a direct constraint on how fast these models can be trained, which changes how teams should size and buy GPU capacity.

    Commercially, Forge places CoreWeave alongside hyperscaler AI platforms such as Amazon Bedrock and SageMaker, Google Vertex AI and Microsoft Azure AI Foundry. Its distinguishing pitch is an open-source foundation and portability across clouds, models and frameworks.

    Inference Is Now on the Training Clock

    In conventional AI development, training and inference are separate phases. A model learns from a fixed dataset, and inference, meaning serving the model to answer requests, comes afterward. Reinforcement learning for agentic models erases that boundary. The model generates attempts at a task, called rollouts. Verifiers check the results and rewards score them, and those scores update the model. Each update produces a new checkpoint, a saved snapshot of the model’s weights, which must be loaded back into the inference servers that produce the next round of attempts. Chowdhary described inference as the bottleneck during rollouts, and the mechanics show why: every reload is time the loop spends waiting.

    That is the problem the 15x figure addresses. CoreWeave RL Rollouts is built on Nvidia’s Dynamo, an open-source framework for distributing inference across many GPUs. It loads new checkpoints into a live deployment, and Chowdhary said the approach improved model reload latency 15-fold against a baseline configuration. The practical effect, in Chowdhary’s framing, is that training keeps moving while inference scales independently. This matters to the people who plan and buy GPU capacity. When inference sits inside the training loop, the size and responsiveness of the inference fleet partly determine training throughput. Teams can no longer budget the two as unrelated line items.

    The claim deserves careful reading. The baseline configuration has not been specified. The figure comes from testing. The capability is in preview. Reload is also only one step in each loop iteration, alongside generating rollouts, scoring them and updating weights, so a 15x gain in reload time does not translate into a 15x gain in end-to-end training speed. CoreWeave has not quantified that overall effect. The direction of the argument is sound. Its magnitude in production is still unproven.

    From GPU Landlord to Platform Vendor

    CoreWeave’s product page lays out Forge as a five-stage loop. Teams run models and agents, observe them through traces in Agent Lens, curate failures into versioned datasets, improve the model through fine-tuning, reinforcement learning or distillation, and evaluate the new version before it ships. Pricing models span per-token serverless inference on open models and dedicated inference, where a customer’s own weights run on GPUs it chooses. Chowdhary described the strategy as building managed services in layers on top of the infrastructure, so that each service builds on the others.

    The economic logic is familiar from the hyperscalers. Raw GPU capacity is easiest to sell when it is scarce, and it is easiest to compare on price when it is not. Software layers give a provider other ways to compete. Customers who keep their experiments, traces, datasets and model registry on one platform have reasons to stay. Per-token pricing lets the provider capture value from efficiency gains on hardware it already owns. The free tier widens the funnel to individual developers who would never sign a capacity contract. Amazon, Google and Microsoft have pursued the same strategy for years by bundling compute with managed AI platforms. Forge is CoreWeave’s entry into that contest.

    CoreWeave’s chosen point of difference is openness. Chowdhary emphasized open-source tools such as vLLM and contributing back to open projects, and the product page promises work that is not locked to one cloud, model or framework. That is a credible counter to proprietary platform stacks. It also creates a tension the company will have to resolve in practice, because an integrated, connected loop is valuable precisely because it is sticky. Portability will be judged by how easily customers can move data, traces and registry entries out, not by the launch language.

    Reading the Inference Demand Signal Carefully

    The demand case rests on one data point. In a theCUBE Research survey of CoreWeave customers and prospects, one healthcare customer’s share of inference workload rose from about 10% in its first year to 40% in its second, with roughly 50% expected within 12 months. That fits the thesis that serving models, not training them, will drive the next phase of AI cloud economics. However, it is a single customer drawn from CoreWeave’s own customer base. It illustrates a trend and does not measure one.

    The infrastructure implication still holds. As more of a GPU’s working life goes to serving requests, providers compete on how much useful output each GPU produces. CoreWeave lists Nvidia’s Hopper, Blackwell and Vera Rubin generations, which competitors can also buy. When the hardware is shared, efficiency comes from the software above it. Chowdhary pointed to that layer: the vLLM serving engine, quantized models, which use lower-precision numbers to shrink a model and speed it up, and custom speculative decoders, in which a small draft model proposes output that the large model verifies. Forge packages that tuning as a product. Over time, it could matter as much as the size of the fleet.

    Background

    CoreWeave began as a GPU-based cryptocurrency operation and pivoted to become a specialized AI cloud, renting large clusters of Nvidia GPUs to AI labs and enterprises. It listed on Nasdaq under the ticker CRWV in 2025 and acquired Weights & Biases, the AI experiment-tracking company, the same year. That tooling now sits inside Forge. Its infrastructure catalog spans Nvidia’s Hopper, Blackwell and Vera Rubin GPUs, managed Kubernetes, AI object storage and fleet-management software.

    Specialized GPU clouds grew on demand for model training, which consumes large blocks of capacity for weeks at a time. As AI products reach users, inference, the continuous work of running models to answer requests, takes up a growing share of compute. That has pushed providers to compete on software efficiency and managed services, the territory where Amazon Web Services, Google Cloud and Microsoft Azure have long operated.

    Sources

    Source: AI inference gets a full stack as CoreWeave launches Forge — SiliconANGLE’s interview with CoreWeave’s Urvashi Chowdhary on Forge, inference optimization and RL Rollouts.

    Primary sources: CoreWeave Forge → Models, agents, training, and inference (CoreWeave product page).

  • Frore’s 35% Tokens-per-Watt Claim Turns GPU Hotspots Into an AI Power Lever

    Frore’s 35% Tokens-per-Watt Claim Turns GPU Hotspots Into an AI Power Lever

    TL;DR · 30-second read

    The Short Version

    A California company called Frore Systems says it can get up to about a third more work out of the chips that run artificial intelligence, without using more electricity.

    The trick is cooling. These chips get extremely hot in a few tiny spots, and when they overheat they slow themselves down. Frore places small pieces of diamond, which moves heat better than metal, right over those spots.

    Why it matters: electricity is now the scarcest ingredient in building artificial intelligence. Squeezing more from the same power could mean fewer new buildings. Frore has not yet published independent test results or named customers.

    Frore Systems announced LiquidJet Diamond on October 8, 2026, in San Jose, California. The product is a liquid-cooling coldplate (the metal block that sits on a chip and carries heat away into circulating fluid) with diamond wafers built into its base, positioned directly over the hottest regions of a GPU. Frore says the design cools hotspots of up to 770 watts per square centimeter and delivers 30–35% more AI tokens per watt than a conventional skived coldplate.

    The company will demonstrate LiquidJet and LiquidJet Diamond cooling a 2,300-watt NVIDIA Rubin GPU, along with its LiquidJet Nexus integrated coldplate system for half-height compute trays, at the OCP Global Summit in San Jose, October 12–16, 2026.

    Executive Summary

    LiquidJet Diamond extends Frore’s existing LiquidJet coldplate, which uses what the company calls a 3D ultra short-loop multi-stage design customized to each GPU’s power map. Frore says the base product lowers die temperature by 12°C, worth 25% more tokens per watt, and the diamond spreader lowers hotspot temperature a further 10°C, lifting the gain to 35%. Against a skived coldplate, it claims over 150% higher maximum hotspot power density, 75% higher heat transfer efficiency, a GPU running 16–22°C cooler, and a drop-in upgrade path.

    The significance lies in the framing rather than the hardware alone. Frore is arguing that in AI data centers capped by grid connections and transformer capacity, chip temperature is a capacity variable: a cooler chip produces more useful output from the same fixed power budget. If the numbers hold up under customer testing, cooling moves from a facilities cost line to a revenue lever.

    The headline claim of up to 35% higher profitability, however, rests on a chain of assumptions about demand, pricing and component cost that Frore has not detailed, and the performance figures have not yet been independently validated.

    Why a Few Square Millimeters Can Set a Data Center’s Output

    A modern AI accelerator does not heat evenly. Some regions of the die, the slab of silicon doing the computation, run much hotter than others. A chip protects itself by slowing down when its hottest point crosses a temperature limit, a behavior known as thermal throttling, and hotter silicon also leaks more electrical current that does no useful work. Both effects mean the temperature of the worst spot on the chip, not the average temperature of the rack, can determine how much computing each watt buys.

    That is the mechanism behind Frore’s numbers. The base LiquidJet design cuts die temperature by 12°C, which Frore equates to 25% more tokens per watt. Tokens are the chunks of text an AI model reads or writes, so tokens per watt is a rough measure of useful AI output per unit of electricity. Adding diamond, which conducts heat far better than copper, directly over the hotspots cuts hotspot temperature a further 10°C and lifts the claimed gain to 35%. Frore says the diamond version handles hotspot power density more than 150% higher than a skived coldplate, a statement about peak heat concentrated in a small area rather than total heat removed.

    The reason that ratio matters is the constraint Frore itself names: utility interconnection (the queue to connect a site to the grid), transformer capacity and on-site power delivery cap what an AI facility can draw. In a building whose power is fixed, more tokens per watt is effectively added capacity without a new substation. Chief executive Seshu Madhavapeddy makes that case directly, saying that with 35% more tokens from the same resources, hyperscalers “would reach their growth goals faster with fewer AI Factories.” The operators most exposed are those deploying the newest, highest-power GPUs, where hotspot density climbs fastest.

    From Tokens per Watt to Profit Takes Several Assumptions

    The release headline promises up to 35% higher profitability; the body says tokens per watt and revenue rise 35%, “with the increased revenue flowing directly to profit.” Each step in that chain carries conditions. Extra tokens only become revenue if there are buyers for them at the same price. Extra throughput only materializes if the GPU is limited by heat rather than by memory bandwidth, networking or software. And the gain only flows straight to profit if the diamond coldplate costs no more over its life than the part it replaces, a figure Frore has not published.

    The temperature figures also mix measures. The 12°C reduction refers to die temperature and the additional 10°C to hotspot temperature; their sum, 22°C, matches the top of the 16–22°C range Frore quotes for a cooler GPU, and the 30–35% tokens-per-watt range suggests results vary by chip and workload. “Up to” marks 35% as the best case. The baseline matters too: “traditional skived coldplates” covers a wide range of designs, and the size of any improvement depends on which one is being compared.

    None of this makes the direction implausible. Cooler silicon generally runs faster and wastes less power, and hotspot-targeted cooling is a logical response to rising chip power. The open question is magnitude under real data center conditions, which is what customer qualification testing exists to establish.

    The Harder Claim Is Manufacturing at Hyperscale

    Frore positions LiquidJet between two incumbent approaches. Skiving, in which thin fins are carved from a copper block, is mature and cheap but, Frore argues, cannot deliver the precision today’s GPUs demand. Electrochemical additive manufacturing (ECAM), a form of 3D metal printing, offers design freedom but, in Frore’s telling, cannot meet hyperscaler quality, reliability, volume or timelines. Frore says it combines semiconductor-fabrication precision with high-volume manufacturing, with production operations in Taiwan.

    Size is part of the pitch. Frore cites die areas ranging from 1,488 mm² for NVIDIA Blackwell Ultra to 46,225 mm² for the Cerebras Wafer-Scale Engine, roughly 30 times larger, and claims its process is the only one able to support the power, performance and size requirements of such chips. Its LiquidJet Nexus system goes further, integrating multiple coldplates and eliminating hoses, connectors and manifolds to enable half-height (½U) compute trays, which would mean fewer potential leak points and denser servers.

    If those claims survive qualification, the pressure falls on suppliers of skived coldplates and on 3D-printing approaches competing for the same high-end sockets. The “seamless drop-in upgrade” claim is commercially important because it lowers switching costs, but hyperscale buyers typically qualify cooling hardware over long cycles, and a component that sits directly on a multi-thousand-dollar GPU faces a high bar for reliability evidence.

    Background

    Frore Systems built its reputation on AirJet, a solid-state active air-cooling chip used in consumer, industrial and Internet-of-Things devices, including a concept ultra-slim notebook unveiled with Lenovo. The company, which describes itself as an AI industry unicorn, has since moved into data center liquid cooling with LiquidJet and LiquidJet Nexus, and earlier published a report, “Chip to Chiller,” arguing that heat costs AI facilities a large share of their compute capacity.

    The broader market context is a shift from air cooling to direct-to-chip liquid cooling as AI accelerators have climbed into the thousands of watts per chip. With grid interconnection and transformer capacity now limiting how fast AI data centers can grow, suppliers across the thermal stack are increasingly pitching their products on output per watt rather than on cooling alone. The OCP Global Summit, run by the Open Compute Project industry consortium, is a key venue where hyperscalers evaluate such hardware.

    Sources

    Source: LiquidJet Diamond Coldplates can increase AI Factory profitability by up to 35% — Frore Systems’ announcement of its diamond-spreader liquid-cooling coldplate for AI GPUs, distributed via PR Newswire.

  • At 40 Minutes per Trailer, Oracle’s AI Gas Stopgap Has a Ceiling

    At 40 Minutes per Trailer, Oracle’s AI Gas Stopgap Has a Ceiling

    TL;DR · 30-second read

    The Short Version

    Oracle, the business-software giant, is building huge computer warehouses for artificial intelligence faster than the gas pipelines meant to power them can be finished. So it is filling trucks with natural gas and driving the fuel to the sites instead.

    It works, but the gas costs roughly four times its normal price. It also does not go far. Running even a small slice of one planned site in New Mexico would need a fresh truckload about every 40 minutes, day and night.

    The lesson is that building artificial intelligence now depends on pipes, permits and trucks, not just computer chips.

    Bloomberg reported that Oracle has been trucking compressed natural gas (CNG) directly to several data center sites so that on-site generators can run before permanent gas pipelines are ready. According to people familiar with the work, truck deliveries kept a data center outside Salt Lake City moving for more than a year while Oracle waited for a pipeline. The same approach is supporting initial work at a campus Oracle is building for OpenAI in Shackelford County, Texas. Oracle is now considering it for Project Jupiter in New Mexico, where a delayed pipeline threatens the timeline of a site with an ultimate footprint of 2.45 gigawatts.

    Certarus, a unit of Superior Plus, supplied the Utah deliveries, and VoltaGrid supplies Texas. Oracle publicly thanked VoltaGrid for “finding cost-effective power solutions to ensure our customers can access OCI capacity on schedule,” referring to Oracle Cloud Infrastructure. Last month Oracle sent the New Mexico project’s developer a force majeure notice. Oracle shares closed down 5.5% at $135.69, their steepest one-day drop since July 16.

    Executive Summary

    The announcement-free headline here is a supply chain story. Many large AI data centers have turned to on-site gas-fired power plants because grid connections take too long. Those plants still need fuel, and fuel normally arrives by pipeline, which can take months to design and build and can be held up by regulators. Oracle is filling that gap with “virtual pipelines”: fleets of trucks hauling compressed gas from a pipeline tap to the site, where it is decompressed and burned in generators.

    The approach buys time, but it has limits. Jack Weixel of East Daley Analytics puts the delivered cost at roughly four times the price of gas at a major pipeline hub once labor, specialized equipment and truck fuel are included. Ellie Holbrook of SemiAnalysis estimates that if Project Jupiter ran just 100 megawatts on trucked gas, which is about 4% of its planned capacity, each large trailer would supply only about 40 minutes of electricity. Trucked gas can energize the first phases of a campus. The date the whole campus comes online is still set by the pipeline.

    For the industry, the episode adds a new item to the list of things that can delay a data center. Developers already track turbines, transformers and interconnection queues. Gas logistics now belongs on that list too, along with the regulators who approve pipeline routes.

    The Trailer Math Sets the Ceiling

    Holbrook’s figures are the clearest way to see how far a virtual pipeline can stretch. At a 100-megawatt load, one large CNG trailer lasts about 40 minutes. A day has 1,440 minutes, so running that load around the clock means roughly 36 trailer loads every day, or one arrival every 40 minutes with no gaps for weather, traffic or a missed fill. Each load is a full cycle. Gas is drawn from a pipeline, compressed into the trailer, driven to the site (sometimes for several hours), then decompressed into the generators. A trip that long ties up several trailers on the road for every one unloading, so the fleet behind a steady stream of arrivals is much larger than the arrival count.

    That 100 megawatts is only about 4% of Project Jupiter’s planned 2.45 gigawatts. As a rough illustration, if fuel demand scaled in a straight line with load, the full footprint would need close to 900 trailer loads a day. No one has proposed doing that, and the arithmetic shows why. Trucked gas is a bridge sized for early phases, such as commissioning, initial server halls and construction power. It cannot replace a pipeline at campus scale.

    The consequence for people planning AI capacity is straightforward. A virtual pipeline can move the date a site first produces revenue. It does not move the date the site reaches full capacity. That second date still depends on Energy Transfer’s rerouted pipeline, which has slipped from this summer to next year.

    Fuel Logistics Joins Turbines and Grid Queues on the Critical Path

    On-site gas generation became popular because it let developers skip the years-long wait to connect to the electric grid. The Oracle sites show that the workaround brings its own bottleneck. A gas plant without a gas pipeline is idle equipment, and pipelines face many of the same constraints as power lines. In New Mexico, state regulators rejected Energy Transfer’s proposed route, forcing a reroute that pushed back the in-service date. The approval delay simply moved from the power line to the pipe.

    This has practical effects on how sites are chosen and planned. When fuel may need to arrive by road for a year or more, as it did in Utah, factors that rarely appeared in data center site studies start to matter. These include the distance to a pipeline tap with available compression, road access able to handle constant heavy-truck traffic, and on-site space for decompression equipment and trailer staging. Developers comparing sites need to weigh the logistics of the interim fuel supply alongside land, water and fiber.

    The suppliers stand to benefit. Until now, CNG trucking has mostly served remote mining and oil drilling operations far from the grid. Certarus and VoltaGrid are now marketing to data center builders as a transitional power source. Certarus said longer timelines for pipeline and grid connections “have created a growing need for rapidly deployable and scalable energy solutions.” Whether these services can scale is a separate question. The number of trucks, compressors and trailers available, along with the drivers to move them, sets a physical limit on how many large campuses can rely on trucked gas at the same time.

    Paying Four Times for Speed

    Oracle described VoltaGrid’s service as “cost-effective,” while East Daley’s estimate puts trucked gas at about four times hub prices. The two claims can both hold, because they measure different things. Per unit of fuel, trucked CNG is expensive. Compared with leaving finished data halls empty, missing delivery dates for customers such as OpenAI, or facing contractual exposure, a costly fuel bill can still be the cheaper choice. Oracle has not disclosed the numbers that would show which effect is larger.

    Those numbers matter because of Oracle’s financial position. Its free cash flow is negative and is expected to stay that way until more of its AI data centers are completed, so each month a site comes online earlier brings revenue forward. The force majeure notice points to the same pressure from another direction. A force majeure clause excuses a party from certain obligations when events beyond its control occur, and Oracle’s notice may shield it from some payments to the New Mexico developer if delays continue. The trucks and the notice are two responses to the same pipeline delay. One keeps work moving, and the other limits the cost if work stalls.

    The open question for investors and customers is where the extra cost lands. It could reduce Oracle’s margins in the early phases, be passed through to the tenant, or be absorbed through the developer agreement. None of these arrangements has been made public, and the answer will determine whether trucked gas is a minor bridging expense or a significant cost of speed.

    Background

    Oracle, long known for database and business software, has become a major builder of AI data centers through Oracle Cloud Infrastructure (OCI). It has committed billions of dollars and its reputation to delivering capacity quickly for customers including OpenAI, and its free cash flow is expected to remain negative until more of those facilities are complete. Its projects include a campus for OpenAI in Shackelford County, Texas, a data center near Salt Lake City, and Project Jupiter in New Mexico.

    Power has become one of the main constraints on the AI data center buildout. Many large developments now generate their own electricity with on-site natural gas plants rather than waiting for grid connections. That approach depends on gas pipelines, which take months to design and build and can be delayed by regulators. Compressed natural gas trucking companies such as Certarus and VoltaGrid, which previously served mainly remote mining and drilling operations, are now offering their services to data center developers as a temporary fuel supply.

    Sources

    Source: Oracle Moves Gas by Trucks to Avoid Data Center Power Delays (SupplyChainBrain, via Bloomberg) on Oracle’s use of trucked compressed natural gas to keep data center projects in Utah, Texas and New Mexico on schedule.

  • Lumentum’s Sellout to 2029 Turns AI Network Optics Into a Long-Lead Item

    Lumentum’s Sellout to 2029 Turns AI Network Optics Into a Long-Lead Item

    TL;DR · 30-second read

    The Short Version

    Lumentum, a company most people have never heard of, makes tiny lasers. They turn computer data into flashes of light so it can travel through glass fiber between the thousands of chips inside the huge buildings that run artificial intelligence.

    Its chief executive says its factories are fully booked into early 2029. For some products, about seven in ten of the parts customers want will still be missing by the end of next year.

    The race to build artificial intelligence is usually told as a story about chips. This shows a small, unglamorous part can set the pace too.

    Lumentum Holdings, the San Jose-based maker of lasers and optical components that carry data inside AI data centers, says its optical component production is fully booked into early 2029. Bloomberg reported on October 8, 2026 that the company, which it described as Nvidia-backed, sees its opto-parts capacity sold out to 2029. GuruFocus reported that chief executive Michael Hurlston said roughly 70% of demand for certain products will remain unmet by the end of next year, and about 30% of demand for other products will go unfilled until 2028. GuruFocus put the company’s market value at about $95 billion.

    The comments came two days after Lumentum filed its definitive proxy statement with the SEC on October 6. In that filing, the company called fiscal 2026, the year ended June 27, 2026, a “landmark year.” It reported record shipments of high-speed transceivers, transport products and laser chips, and said it had started production of next-generation 1.6T modules. It also said its optical circuit switch business had passed a $100 million quarterly revenue run rate.

    Executive Summary

    Lumentum is telling the market that it cannot make enough of the lasers and optical modules that link AI accelerators together. Its production is committed into early 2029. For some product lines, most of the demand it sees for next year will go unserved. Those are the parts that convert electrical signals to light and back again, so that thousands of GPUs can behave like one machine.

    The significance for infrastructure builders is timing. AI capacity plans have mostly been framed around securing GPUs, power and buildings. A multi-year sellout at a major laser supplier adds optical interconnect to the list of items that have to be reserved well before a cluster is commissioned. It also matters because Lumentum’s October 6 proxy statement argues that optics will move deeper into the server rack, which would raise the amount of optics each AI system needs.

    What is substantiated is narrower than the headline figure suggests. The company has stated a booking horizon and percentages of unmet demand. It has not disclosed which products are affected, who holds the committed capacity, or whether those commitments are binding orders or allocations.

    Why a Laser Sellout Becomes a Cluster Schedule

    Modern AI training and inference clusters are networks first. Every server in a scale-out cluster, meaning the fabric that connects many servers across a data center hall, reaches its switches through optical transceivers. A transceiver is a plug-in module that turns electrical data into light for the fiber run and back again. Each one depends on laser chips. Lumentum makes both the chips, including EML (electro-absorption modulated) and CW (continuous-wave) lasers grown on indium phosphide wafers, and finished modules such as the 1.6T (1.6 terabits per second) parts it began producing in fiscal 2026. A rack of GPUs that arrives without its optics cannot join the cluster, however many accelerators it holds.

    That is the mechanism behind the long-lead label. If Lumentum’s production is booked into early 2029, a buyer placing a new order with Lumentum today is effectively scheduling delivery more than two years out. Hurlston’s figures sharpen the point. Roughly 70% of demand for certain products will still be unmet by the end of next year, and about 30% of demand for other products will be unfilled until 2028. For hyperscalers, neocloud operators and the developers whose tenants need a working network fabric on a fixed go-live date, optics now has to be secured early, the way developers already secure grid capacity and electrical equipment.

    The claim has limits that matter. Lumentum is one supplier, and it has not said what share of the market its booked capacity represents. Buyers with diversified supply may be able to source modules or lasers elsewhere, if other vendors have room. The defensible conclusion is that optical interconnect has become a dated procurement constraint for anyone depending on Lumentum’s capacity. It is not proof that every AI cluster will be delayed, and it does not mean GPUs are no longer scarce.

    Scale-Up Could Raise the Optics Count Per Rack

    The proxy statement explains where Lumentum thinks demand goes next. Its results to date came from scale-out networks and scale-across networks, which link separate data center sites. The company says it is most excited about scale-up, the tight connections between GPUs inside a single rack or system. Today those links are largely electrical, over copper. Lumentum describes a shift toward optical connectivity within the rack that “has the potential” to drive its performance for years.

    If that shift happens at scale, the number of lasers needed per GPU rises. And it would do so at a moment when Lumentum says its existing capacity is already spoken for. The company also points to co-packaged optics (CPO) and near-packaged optics (NPO), designs that place the optical engine on or next to the switch or processor package. Lumentum says these designs eliminate the need for a digital signal processor (DSP), cutting power use and cost. They also depend on the ultra-high-power lasers Lumentum is developing. The scale-up thesis is the company’s forecast rather than a disclosed order book. Still, it explains why a supplier might describe demand in years rather than quarters.

    Optical Circuit Switches: A Reported Number and a Forecast

    Lumentum’s optical circuit switch (OCS) line is the clearest growth figure in the filing. An optical circuit switch redirects beams of light between fibers without converting the signal to electrical form, which can save power in large AI networks. The company says the platform has passed a $100 million quarterly revenue run rate. That is a reported operating figure.

    The next claim is of a different kind. Lumentum says it expects the product line to generate “revenue in the billions in upcoming years,” citing technological differentiation and a first-mover advantage. That is a management expectation with no stated timeline or customer base attached. Readers should weigh it as guidance, not as booked business. The same distinction applies to the 2029 sellout. A booking horizon is a strong demand signal, but its value to the company depends on the pricing and terms attached to it, and Lumentum has not disclosed either.

    Who Gains and Who Waits

    A supplier with sold-out capacity usually has leverage on price and terms. That can support margins unless long-term agreements locked in earlier pricing. Customers who committed early hold an advantage over those arriving now. In practice, AI operators that reserved optics alongside their GPU allocations are better placed than late entrants or smaller cloud providers. Module assemblers that buy laser chips rather than make them are also exposed to whatever allocation decisions upstream laser suppliers make.

    For data center developers, the practical effect is coordination. A powered, cooled hall delivers no revenue until the tenant’s network is live, and that network now has its own multi-year lead time. Lumentum says it has expanded wafer capacity across its global indium phosphide fabs and qualified both CW and EML process flows. It has not said how much capacity that adds or when the backlog eases.

    Background

    Lumentum was spun off from JDS Uniphase in 2015 and is headquartered in San Jose, California. It makes lasers, optical transceivers, transport products and, more recently, optical circuit switches. Its cloud and networking business supplies the components that move data between servers and switches in large data centers. The company’s fiscal year ends in late June, and fiscal 2026 closed on June 27, 2026.

    Demand for optics has risen with the size of AI clusters. Training a large model means synchronizing thousands of accelerators, and every link between them needs bandwidth. Network speeds have stepped up from 400-gigabit to 800-gigabit and now 1.6-terabit modules. Many of those modules rely on laser chips made from indium phosphide, a semiconductor material produced in specialized fabs that take time and capital to expand.

    Sources

    Source: Lumentum Holdings Inc. Faces Soaring Demand for AI Data Center Components (GuruFocus), which reported CEO Michael Hurlston’s remarks that optical component production is booked into early 2029. Also: Nvidia-Backed Lumentum Sees Opto-Parts Capacity Sold Out to 2029 (Bloomberg).

    Primary sources: Lumentum Holdings Inc., DEF 14A Definitive Proxy Statement, filed October 6, 2026 (SEC EDGAR).

  • Bitdeer’s 67MW Malaysia AI Bet Rests on a 9.5MW Sellout — and No Tenant Yet

    Bitdeer’s 67MW Malaysia AI Bet Rests on a 9.5MW Sellout — and No Tenant Yet

    TL;DR · 30-second read

    The Short Version

    Bitdeer, a company best known for bitcoin mining, is adding a large computer center in Malaysia built for artificial intelligence work. It will draw about as much electricity as a sizable town and is due to be ready by mid-2027.

    The catch: no customer has signed up to rent it yet. Its main proof that people want this is a Malaysian site one-seventh the size, which sold out this summer.

    So it is a bet that customers will come, made while the company pays as much as 13% a year on some of its loans.

    Bitdeer Technologies (Nasdaq: BTDR), through its Bitdeer AI unit, has secured a 67MW AI data center campus in Malaysia, with delivery targeted for the second quarter of 2027, according to MarketScreener, StreetInsider and Crypto Briefing. TheEnergyMag reported that the addition brings Bitdeer’s total to 273.5MW. FourWeekMBA reported that no offtake has been signed. An offtake is a customer’s commitment to take the capacity. Investing.com put the facility at 60MW rather than 67MW. StreetInsider and Stocktwits reported that Bitdeer’s AI infrastructure pipeline tops $10 billion.

    Separately, Stock Titan reported that Bitdeer told readers to disregard a release that had cited a facility target of up to 350MW. MarketScreener reported that alerts on Bitdeer AI’s data center services agreement were withdrawn. The announcement follows a 6-K filed September 29, in which Bitdeer disclosed that its roughly 9.5MW A102 data center in Malaysia had sold out on five-year offtake commitments.

    Executive Summary

    Bitdeer is adding 67MW of AI data center capacity in Malaysia, a country where it already operates. Delivery is targeted for Q2 2027, only a few quarters away, and no customer has committed to the space. The only Malaysian AI contract the company has put on the record is the 9.5MW A102 sellout disclosed in its September 29 filing. The new campus is therefore about seven times larger than the demand Bitdeer has demonstrated in that market.

    The contrast with Bitdeer’s Norwegian project is sharp. At Tydal, the company signed a named tenant at a disclosed price before building, with roughly $1.3 billion of bank-arranged credit support behind the tenant. The Malaysian campus so far has a size and a date, but no tenant, price or financing disclosed. This is a speculative build, meaning one put up ahead of signed demand. It comes as Bitdeer carries collateralized loans at a 13.0% effective interest rate, with $373.5 million of principal falling due within a year of June 30, 2026.

    Two figure discrepancies need clearing up. One is the withdrawn reference to an up-to-350MW target. The other is the 60MW versus 67MW sizing. Until Bitdeer clarifies them, it is hard to say exactly what is being built.

    A 9.5MW Sellout Is the Demand Case for 67MW

    Bitdeer has put only one piece of Malaysian AI demand on the record, and it is small. In its September 29 6-K, the company said Bitdeer AI sold out approximately 9.5MW at its A102 data center in August, with five-year offtake commitments for its AI Cloud business, which sells computing capacity for AI workloads. The new campus, at 67MW, is about seven times that size. It is being added before any customer has committed to it.

    That ratio is the core of the bet. The A102 sellout shows that buyers in Malaysia will sign multi-year commitments for Bitdeer’s capacity. It does not show that buyers exist for seven times as much, on Bitdeer’s timeline and at the price it needs. The calendar is also short. Q2 2027 is two to three quarters away, so Bitdeer has quarters, not years, to sign tenants. If the campus is delivered before it is contracted, Bitdeer carries the cost of a finished facility with no revenue against it until the space fills.

    Different groups are affected in different ways. For AI companies looking for capacity in Southeast Asia, an uncommitted campus with a near-term delivery date gives them room to negotiate. For Bitdeer shareholders, the value of the project depends on when the first tenant signs and on what terms. A further question is how the space will be used. Bitdeer could fill the campus with its own AI Cloud service, as at A102, which means buying its own processors on top of the facility. Or it could lease the space as colocation, where the tenant brings its own hardware. Those two paths carry very different capital requirements.

    Tydal Is the Template Malaysia Has Not Yet Matched

    Bitdeer’s own filings show what a fully contracted AI campus looks like. Under the Tydal, Norway agreement with Volta Tydal AS, described in the September 29 6-K, Bitdeer’s subsidiary will deliver 121 IT MW of contracted critical load, supported by an estimated 133 gross MW. IT MW is the power available to the computers themselves. Gross MW also covers cooling and other building systems. The contract is worth approximately $4.7 billion over a 16-year base term, at an average rate of about $202 per kilowatt per month. Electricity is fully reimbursed by the tenant, and payments rise 3% a year. Bitdeer expects about $2.4 million of annual revenue per IT MW. It also expects a net operating income margin of roughly 90%, meaning the share of revenue left after operating the property.

    The tenant’s obligations are expected to be backed by letters of credit totaling about $1.3 billion, arranged by affiliates of J.P. Morgan and another major financial institution. These are bank guarantees that pay out if the tenant cannot. That backing is subject to customary conditions, and the tenant can terminate without a fee at year 10. So even Bitdeer’s strongest contract has conditions attached. Still, it has a named counterparty, a price, a term and credit support, all disclosed before construction completes. The design-and-construction contract for Tydal is phased to begin in December 2026.

    The Malaysian campus has none of those elements on the record yet. That is not unusual for a project announced early. But it means the 67MW should be read as capacity Bitdeer intends to sell, not capacity it has sold.

    13% Debt Puts a Clock on the Lease-Up

    Bitdeer’s balance sheet makes the timing of tenant signings more important. According to the financial statements filed September 29, the company’s collateralized loans from BIT Group carried a 13.0% effective interest rate in the first half of 2026. Interest expense on those loans was $32.7 million, up from $3.0 million a year earlier. As of June 30, $373.5 million of that principal was due within twelve months, and $142.1 million was due after that. The loans are secured by mining rigs, inventories and data center assets.

    Bitdeer has other obligations as well. It has a $26.0 million bank facility at 10.31% that matures in December 2026. It has $375.0 million of 5.00% convertible notes due 2032, issued in February. It also had roughly 3,669 Bitcoin outstanding under a separate BIT Group Bitcoin loan as of September 29. In August, the company paid about $100 million in cash for roughly 200 acres in Milam County, Texas. On October 8, its board authorized a new $100 million share buyback after fully using a previous $40 million program.

    None of that is unusual for a company expanding quickly. It does mean the Malaysian campus competes for capital with debt repayments, land acquisition and buybacks. Every quarter the campus sits unleased, capital stays committed with nothing coming back. Bitdeer has not said how the campus is being financed.

    One Figure Withdrawn, Another Disputed

    The withdrawn up-to-350MW reference is worth noting in a measured way. Companies correct releases for many reasons, and the figure that stands is the 67MW campus. But Bitdeer has not explained what the 350MW target referred to or whether a larger expansion is still under consideration. For a company that investors increasingly value on its pipeline of planned capacity, that clarification matters.

    The 60MW versus 67MW gap may have a simple explanation. The ratio is close to the IT-versus-gross distinction Bitdeer itself draws at Tydal, where 121 IT MW sits within 133 gross MW. But Bitdeer has not confirmed that this is the reason. It is a question the company can answer quickly. Buyers pricing capacity and investors modeling revenue need to know which number is the sellable one.

    Background

    Bitdeer Technologies Group is incorporated in the Cayman Islands, headquartered in Singapore and listed on the Nasdaq under the ticker BTDR. It built its business on bitcoin mining. Its revenue lines include self-mining, cloud hash rate (selling mining capacity to others), hosting and mining rig sales, plus a growing AI Cloud business. Its financial statements break out figures for Ethiopia, Canada, Malaysia, Norway, Bhutan, the United States and Singapore. The company cites about 3.0 GW of total power capacity across its global site portfolio.

    Like several miners, Bitdeer is repositioning its power-connected sites toward AI computing. In Norway, it is converting its Tydal data center into an AI facility, under a cost-plus construction contract phased to begin in December 2026. In Texas, it bought about 200 acres near its Rockdale facility in August for roughly $100 million in cash, giving it 742MW of existing and pipeline power capacity in Milam County. In Malaysia, its 9.5MW A102 site sold out to AI Cloud customers in August 2026.

    Sources

    Source: Bitdeer Technologies Expands Malaysian AI Data Center Capacity With 67MW Campus (MarketScreener): Bitdeer AI adds a 67MW Malaysian AI data center campus.

    Primary sources: Bitdeer 6-K filed October 8, 2026 (US$100 million share repurchase program); Bitdeer 6-K filed October 2, 2026 (resignation of director); Bitdeer 6-K filed September 29, 2026 (interim report); Exhibit 99.1 to 6-K filed September 29, 2026 (unaudited consolidated financial statements); Exhibit 99.2 to 6-K filed September 29, 2026 (recent developments).

    Further reporting: FourWeekMBA; StreetInsider; Stocktwits; TheEnergyMag; Crypto Briefing; MarketScreener (withdrawn alerts); Stock Titan; Investing.com.

  • EMCOR’s 43.9% Backlog Jump Shows AI Data Center Builds Hinge on Contractor Throughput

    EMCOR’s 43.9% Backlog Jump Shows AI Data Center Builds Hinge on Contractor Throughput

    TL;DR · 30-second read

    The Short Version

    The giant computer warehouses being built for artificial intelligence need armies of electricians, pipefitters and steelworkers. The companies that supply them are signing jobs faster than they can finish them.

    EMCOR, one of the biggest, has about $17 billion of signed work waiting. That is up 44% in a year, while its sales grew about 20%.

    A smaller rival, IES Holdings, just spent $691 million, mostly borrowed, to buy a steel builder with about 4,000 workers. In this boom, the hard part is not winning work. It is having the people and workshops to do it.

    IES Holdings (NASDAQ: IESC) closed its acquisition of DBM Global on October 5, 2026, for total consideration of about $691 million: roughly $545 million in cash and 430,974 IES shares valued at about $146 million. In an 8-K filed October 6, IES said it funded the cash with cash on hand and $525 million of borrowings under a credit agreement it amended the same day to raise borrowing capacity to $700 million. DBM, a structural steel fabricator and erector with about 4,000 employees, generated about $1.5 billion of revenue in the twelve months ended June 30, 2026.

    The deal lands as data center construction drives growth across specialty contracting. Insider Monkey reported that EMCOR Group (NYSE: EME), in its July 30 second-quarter results, posted revenue of $5.15 billion, up 19.8%, while its remaining performance obligations—signed work not yet performed—rose 43.9% to $17.14 billion. IES’s fiscal third-quarter revenue rose 40% to $1.24 billion, led by a 51% jump in its data center–driven Communications segment.

    Executive Summary

    IES Holdings has made its largest acquisition to date, paying about $691 million for DBM Global, which IES describes as one of the largest independent structural steel fabrication and erection platforms in the US. The deal turns an electrical and technology-systems contractor into one that can also engineer, fabricate and erect a building’s steel frame, and IES says it will offer the combination to data center, industrial and infrastructure customers.

    The timing matters because the sector’s order books are expanding faster than its output. EMCOR, the larger rival, grew its backlog 43.9% against 19.8% revenue growth. When signed work outpaces completed work, the pace of data center construction depends increasingly on the contractor side—the skilled crews, fabrication shops and project managers who turn contracts into finished buildings. IES’s purchase brings in about 4,000 employees and more than 2 million square feet of facilities, financed largely with $525 million of bank borrowing.

    For investors, the two companies now frame a choice: EMCOR’s demonstrated conversion of demand into profit at a lower earnings multiple, or IES’s faster growth, which now carries integration and debt-repayment tests.

    EMCOR’s Order Book Is Growing Twice as Fast as Its Revenue

    In EMCOR’s July 30 results, quarterly revenue rose 19.8% to $5.15 billion, while remaining performance obligations rose 43.9% to $17.14 billion. For a contractor, revenue largely reflects work actually put in place during the period; remaining performance obligations are contracted work still to be done. When the second grows more than twice as fast as the first, signed work is accumulating faster than crews, shops and project teams are converting it. The arithmetic makes the point concrete: at the second-quarter pace, $17.14 billion equals about 3.3 quarters of revenue. The year-earlier figures implied by those growth rates put the ratio near 2.8 quarters.

    There are two fair readings of that gap. Larger, multi-year campus contracts naturally lengthen a backlog, and the $17.14 billion is a company-wide figure spanning all of EMCOR’s markets, not data centers alone. But both readings lead to the same operational conclusion: once work is booked well ahead of output, the speed at which buildings get finished is set by contractor throughput—skilled labor, procurement and project management—rather than by the availability of contracts. EMCOR has so far converted efficiently. Earnings per share rose 34.8% to $9.06, faster than revenue, operating income reached $547.3 million at a 10.6% margin, and management raised full-year guidance to $32–$33.25 per share.

    The consequences run in several directions. For data center developers and cloud operators, the contractor’s calendar becomes part of the project timeline alongside power and equipment. For contractors, a backlog becomes profit only if costs stay within contract assumptions, and a larger book carried by the same workforce raises execution risk. For investors, the number to watch is conversion—margin on work completed—rather than the volume booked.

    What IES’s $691 Million Actually Buys

    Read against that backdrop, the DBM Global acquisition is a purchase of delivery capacity. IES’s announcement lists about 4,000 employees, more than 2 million square feet of fabrication and operating facilities, and brands including Schuff Steel, Banker Steel, GrayWolf, DBM Vircon and Aitken. On the disclosed figures, IES paid roughly 0.46 times DBM’s trailing $1.5 billion of revenue, or about $173,000 per employee. What that buys is an operating platform—trained crews, shops and customer relationships that DBM CEO Rustin Roach says in many cases go back decades—that cannot be assembled quickly from scratch.

    The strategic logic is a second trade on the same job site. A data center needs a structural steel frame and dense electrical systems, and IES CEO Matt Simmes said the combined company will offer a broader range of products and services, particularly in data center, industrial and infrastructure markets. A contractor that can deliver both can compete for a larger share of each project’s construction budget. Whether customers prefer to buy the two trades as a bundle is not yet demonstrated, and that is the commercial test the acquisition sets.

    The financing is equally telling. A Wells Fargo-led bank group expanded IES’s facility to $700 million—a $200 million term loan and a $500 million revolver—and $525 million of borrowings covered most of the cash. In this deal, credit was the readily available input. The scarcer one was a structural steel business already operating at scale, and IES paid a full price, in cash and shares, to own it rather than build it.

    Two Ways to Price the Same Build-Out

    At October 5 prices, consensus forward earnings multiples were about 22 times for EMCOR and 27 times for IES—a forward multiple being the share price divided by expected earnings per share. At about $779, EMCOR traded near 24 times the midpoint of its raised guidance. The case for EMCOR rests on operating leverage: earnings growing faster than revenue suggests added work is lifting profitability, not just volume.

    The case for IES rests on its data center exposure. Communications revenue, supported primarily by data centers, rose 51% to $453.1 million with $83.6 million of operating income—a segment margin of about 18%. Consolidated growth is less clean: Infrastructure Solutions benefited from the addition of Gulf Island, and Residential revenue fell 6%. DBM now adds roughly $1.5 billion of structural steel revenue whose margins IES has yet to report, and the company has moved from a debt-free balance sheet in June to $525 million of acquisition borrowing. Executive Chairman Jeff Gendell said IES expects substantial cash flow from both businesses to allow it to rapidly repay those borrowings.

    The risks are mirror images. EMCOR’s is extrapolating unusually favorable execution across a backlog that is growing faster than its revenue. IES’s is integration plus leverage, layered on the same execution demands. Both ultimately depend on the variable the backlog numbers highlight: the ability to deliver the work already won.

    Background

    EMCOR Group is a large US specialty contractor whose electrical and mechanical work—power distribution, cooling and piping—sits inside the buildings it serves, including data centers. Specialty contractors are hired by developers and general contractors to install the systems that make a building function. In AI data centers, more powerful computing racks change not just the equipment bill but the electrical and mechanical installation work surrounding them, which is why these firms have become direct beneficiaries of the build-out.

    IES Holdings, based in the Houston area, designs and installs integrated electrical and technology systems and provides infrastructure services; with DBM it now also engineers, fabricates and erects structural steel, and it has more than 16,000 employees. DBM Global was majority-owned by INNOVATE Corp. before the sale. IES signed the transaction agreement on August 7, 2026, and closed on October 5.

    Sources

    Source: EMCOR vs. IES Holdings: Does a $691 Million Deal Change the Better Buy? (Insider Monkey) — a comparison of the two contractors’ data center-driven growth, backlog and valuations after IES closed its DBM Global acquisition.

    Primary sources: IES Holdings Form 8-K filed October 6, 2026; IES Holdings press release, “IES Holdings Completes Acquisition of DBM Global” (Exhibit 99.1).