Chips & Supply · United States

The AI Boom Has a Gigawatt Accounting Problem

Credited to EE Times · eetimes.com · 20h ago

Photograph · EE Times

Track energized compute, not gigawatts, as AI data centers face delays in memory, networking, cooling, and power.

Broadcom can lend Anthropic up to $42 billion to help finance computing infrastructure. Nvidia says it intends to invest as much as $100 billion in OpenAI as Nvidia systems are deployed. AMD has issued warrants to major customers whose vesting is tied partly to gigawatt-scale GPU purchases.

Capital is clearly available for AI infrastructure. The harder question is what those commitments will produce in the physical world.

The industry increasingly describes AI infrastructure in gigawatts. OpenAI and Nvidia have announced plans for at least 10 GW of Nvidia systems. OpenAI has a separate definitive agreement with AMD covering 6 GW of GPUs and another collaboration with Broadcom covering 10 GW of OpenAI-designed accelerators and networking systems. Those figures look comparable, but they describe different things.

A gigawatt of contracted GPUs is not necessarily a gigawatt of data center capacity. Contracted cloud compute is not the same as installed hardware, and infrastructure described as secured is not necessarily infrastructure that has memory, networking, cooling, and power available to operate.

That distinction matters because AI infrastructure is becoming a systems problem, not simply a semiconductor capacity problem.

I have followed semiconductor capacity cycles for more than two decades, including during my years at Intel. Purchase orders are only one part of the capacity equation. Manufacturing, packaging, memory, and the infrastructure around the chip determine how quickly demand turns into usable systems.

Nvidia’s Vera Rubin reference design illustrates the scale involved. Its published configuration for a 100 MW AI factory includes 40,000 Rubin GPUs and 12 petabytes of HBM4. A simple linear scaling of that reference design would imply roughly 400,000 GPUs and 120 petabytes of HBM4 per gigawatt. That is an illustration, not a universal conversion factor for AI capacity.

The memory requirement alone is significant. Samsung expects HBM to account for nearly 30% of industry DRAM wafer capacity in 2027, up from about 20% currently. Samsung also notes that HBM and standard DRAM share wafer production capacity, so growth in one affects the capacity available for the other.

The next challenge is connecting the processors. Large AI clusters require switches, network-interface cards and large volumes of high-speed optical links. At rack level, rising power density adds liquid cooling, coolant distribution, large power systems, and more complex commissioning.

The GPU is therefore only one component in a much larger industrial system. The supply chain extends from semiconductor manufacturers and memory producers to cooling suppliers, electrical contractors, utilities, and power generation companies.

That becomes clear when construction starts.

JLL estimates that a 50 MW data center takes about 18 months to build globally. Developers are ordering some equipment as much as two years in advance, yet 57% of projects experienced construction delays of at least three months in 2025. In the U.S., JLL puts average equipment lead times at roughly 42 weeks, 83% above 2019 levels. Transformers and switchgear average about 43 weeks, while generators average roughly 51 weeks. A contracted gigawatt can therefore exist long before it becomes usable compute.

The difference also becomes more important closer to the power grid. Semiconductor capacity is relatively fungible. Wafers can be reallocated between customers, and some production steps can be moved between suppliers. A substation, transmission connection, or gas pipeline belongs to a specific location.

That makes power increasingly central to AI deployment. Enverus forecasts that 22.5 GW of new U.S. data center demand between 2026 and 2030 will be served by behind-the-meter generation, equivalent to about 36% of projected U.S. data center capacity additions during that period. Its forecast includes 29.6 GW of gas-fired generation.

Developers are effectively considering power generation as part of the data center project rather than waiting for conventional grid expansion. That can shorten some timelines, but it creates another set of constraints around turbines, engines, transformers, pipelines, and permits. Enverus itself cautions that access to a theoretical power source does not automatically make a site ready for construction.

The financing structures now emerging around AI make these physical constraints more relevant.

OpenAI and Nvidia’s announced partnership calls for at least 10 GW of Nvidia systems, representing millions of GPUs. Nvidia says it intends to invest up to $100 billion in OpenAI as the systems are deployed. The first gigawatt is targeted for the second half of 2026 using Vera Rubin. Importantly, the arrangement was announced as a letter of intent rather than a completed purchase obligation.

OpenAI separately entered a definitive agreement with AMD covering 6 GW of GPUs across multiple generations. The first 1 GW deployment using MI450-series products is scheduled to begin in the second half of 2026. AMD also issued OpenAI a warrant for up to 160 million AMD shares, with vesting linked in part to GPU-purchase milestones.

A separate OpenAI collaboration with Broadcom covers 10 GW of OpenAI-designed accelerators and networking systems, with deployment targeted to begin in the second half of 2026 and finish by the end of 2029.

Those announcements cannot simply be added together and treated as separate data center capacity. They describe different layers of the same infrastructure buildout.

The Meta-AMD agreement shows why that distinction matters. AMD describes the broader relationship as covering up to 6 GW, but its SEC filing says Meta made a binding commitment for the initial 1 GW. Full vesting of Meta’s warrant for up to 160 million AMD shares depends on purchases eventually reaching 6 GW. The headline number and the current binding obligation are not identical.

Anthropic provides another example of how financing and hardware are becoming intertwined.

According to Anthropic’s IPO prospectus, as reported by Reuters, Broadcom agreed to lend the company up to $42 billion to help finance infrastructure. Reuters reported that the facility could cover roughly one-third of a five-year, $125.2 billion commitment for TPU computing capacity. Anthropic is also expected to become Broadcom’s largest chip design customer in 2027.

This does not mean the underlying demand is artificial. It does mean that the roles of supplier, customer, investor, and financier are becoming increasingly difficult to separate.

The same future unit of compute can now appear in several places: as a semiconductor order, a cloud compute commitment, a data center project, a financing obligation, and eventually as capacity described by an AI company as secured.

Each disclosure may be accurate on its own. Problems arise when they are added together as though each represents separate physical capacity. That is the accounting issue behind the industry’s growing use of gigawatts. A gigawatt is a precise unit of electrical power. It is much less precise as a measure of AI capacity.

None of this means the AI infrastructure boom cannot be built. Semiconductor manufacturers are expanding production, memory suppliers are adding HBM capacity, and data center developers are ordering equipment earlier and looking for alternatives to conventional grid connections.

The more important constraint may be timing. Capital can be committed quickly. A GPU agreement can be signed years before delivery, and a cloud customer can contract for compute before the underlying facility exists. Usable compute becomes available only when the semiconductor, memory, network, rack, cooling, and power layers are ready in the same place.

A one-year delay does more than move a construction schedule. It affects utilization, financing costs, depreciation assumptions, and the timing of revenue throughout the infrastructure supply chain.

For that reason, announced gigawatts may become less useful as the AI buildout matures. A better measure may be time to energized compute: How long it takes a financial commitment to become a functioning system that can actually run workloads.

The AI industry has shown that it can finance enormous amounts of future capacity. The next test is how quickly the physical infrastructure can catch up.

AI Power Demands Push GaN into Data Center Design

AI Data Centers Make Power, Cooling Critical to Scaling

The Energy Barrier Reshaping AI Hardware

OpenAI and Nvidia Systems Partnership Supports the announced 10 GW framework, millions of GPUs, intended investment of up to $100 billion and the initial Vera Rubin deployment. https://openai.com/index/openai-nvidia-systems-partnership/

OpenAI and AMD Strategic Partnership Supports the 6 GW definitive agreement, first 1 GW MI450-series deployment and warrant for up to 160 million AMD shares. https://openai.com/index/openai-amd-strategic-partnership/

OpenAI and Broadcom Strategic Collaboration Supports the 10 GW accelerator and networking collaboration and targeted 2026-2029 deployment period. https://openai.com/index/openai-and-broadcom-announce-strategic-collaboration/

Nvidia Vera Rubin NVL72 Supports the 100 MW reference configuration with 40,000 Rubin GPUs and 12 PB of HBM4. The 1 GW figures in the article are derived by simple linear scaling for illustration only. https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/

AMD SEC filing, Meta Agreement Supports the binding initial 1 GW commitment, broader arrangement up to 6 GW and warrant for up to 160 million AMD shares. https://www.sec.gov/Archives/edgar/data/2488/000000248826000045/amd-20260223.htm

Reuters, Broadcom, and Anthropic Supports the financing facility of up to $42 billion, the $125.2 billion five-year TPU compute commitment and expected 2027 customer relationship. Reuters

Reuters, Samsung HBM capacity Supports Samsung’s expectation that HBM will approach 30% of industry DRAM wafer capacity in 2027, versus about 20% currently. Reuters

JLL, Global Data Center Outlook 2026 Supports construction timelines, 2025 project delays and U.S. equipment lead-time estimates. https://www.jll.com/content/dam/jllcom/en/global/documents/reports/research-reports/26-research-global-data-center-outlook-new.pdf

Enverus, behind-the-meter generation forecast Supports the forecast for 22.5 GW of U.S. data center demand served behind the meter, the 36% share of capacity additions and the modeled gas-generation requirement. https://www.enverus.com/newsroom/behind-the-meter-generation-forecast-skipping-the-queue/

Enverus, site readiness Supports the discussion of pipeline, permitting and infrastructure constraints on data center sites. https://www.enverus.com/blog/data-center-sites-powered-by-gas-isnt-the-same-as-ready-to-build/

RELATED TOPICS: AI AND BIG DATA, POWER MANAGEMENT

Ron Honig is co-CEO of From Honig Family Office. He previously spent more than two decades in the technology industry, including at Intel, with a long-standing focus on the semiconductor sector.

Chips & SupplyUnited States

Ultra-Compact Photonic Bends Reduce Loss While Meeting Foundry Rules (Georgia Tech)

Credited to Semiconductor Engineering · semiengineering.com20h ago

Useful

Researchers at Georgia Institute of Technology published a technical paper titled “Low-loss ultra-compact photonic bends with enhanced fabrication tolerance for commercial foundry processes.”

Abstract Excerpt: “We present a methodology and experimentally validate it with the design of high-performance integrated photonic waveguide bends in extremely compact footprints (2.3 µm radius). The approach uses what we believe to be a novel contour constraint within a topology optimization framework to improve performance and reduce sensitivity to fabrication variability while ensuring compliance with commercial foundry design rule checks.”

Find the technical paper here. September 2026. Wong, Joshua J., Robert P. Pesch, Archana Kaushalram, Jacob M. Hiesener, and Stephen E. Ralph. “Low-loss ultra-compact photonic bends with enhanced fabrication tolerance for commercial foundry processes.” Optics Letters 51, no. 18 (2026): 5185–5188. https://doi.org/10.1364/OL.611581

Photonics Fundamentals For Electronics Engineers: eBook

Original · Semiconductor Engineering

Front Method

Chips & Supply · United States

Meta Muse popularity lifts AMD stock to fresh highs as AI agents juice CPU sales

Credited to CNBC · cnbc.com · 22h ago

The personal AI agent boom is giving Advanced Micro Devices and Intel new life after the companies spent years watching Nvidia and its GPUs dominate the artificial intelligence market.

The central processing unit, or CPU, was once the primary workhorse of a computer server before Nvidia's graphics processing units took center stage with the launch of ChatGPT and the start of the generative AI era. Now CPUs are back, lifting the fortunes of market leaders AMD and Intel, whose chips specialize in powering long-running agents that can work for hours on their own.

Last week OpenAI released its AI agent Dots, following the lead of Meta, which debuted Muse in early September. Muse caught on so quickly that it topped the Apple App Store in less than two weeks. Well before that, AMD and Intel were already huge winners in the stock market this year due largely to surging demand for CPUs, but they've picked up steam in the past month, jumping 32% and 21%, respectively, as of Monday's close beating all of tech's megacaps.

The rally continued on Tuesday, with AMD jumping almost 3% to a fresh high of $649.42. AMD's growing strength in both CPUs and GPUs has pushed the company into the trillion-dollar market cap club.

"As more agents are created and developed, and more people start to use them for more tasks, it's going to start to shift the workload away from GPUs and onto CPUs," said Ryan Shrout, president of Signal65, which consults on AI hardware and costs.

Users learned about the power behind the new agents by asking them. Muse tells users that it runs on an AMD-powered computer, while Dots said it runs on a virtual computer powered by AMD's EPYC-branded CPU.

A Meta spokesperson told CNBC that the company designed its system to use whatever kind of CPU is available.

"We take a diverse approach to our hardware and are largely CPU-agnostic by design, which gives us the most flexibility in acquiring capacity," the Meta spokesperson said in a statement.

OpenAI similarly uses multiple CPU providers, meaning AMD isn't the only game in town.

Most servers at hyperscalers — Amazon, Google, Meta and Microsoft — use CPUs made by Intel or AMD. But nearly every major cloud provider is also developing custom chips of their own, typically based on Arm technology. Arm announced its own CPU for agents in March, with Meta as the debut customer, and its stock price has more than doubled since.

Earlier this year, Nvidia released Vera — a fully redesigned central processor — along with an entire rack filled only with the CPUs. The chip giant says it built the new Vera CPUs specifically for agents and expects CPUs to be a $200 billion market by 2030.

But it's one part of the chip industry where Nvidia isn't currently the favorite, analysts say.

"It's just a new product and they're going to be competing against a lot of different players," Jordan Klein, an analyst at Mizuho Securities, said in an interview, adding that he believes AMD currently has the "best product."

CPUs are driving a healthy portion of AMD's growth. Revenue in the company's data center business more than doubled to $6.7 billion in the quarter ended in June, accounting for almost 60% of total sales.

" The conversation has changed quite a bit over the last eight to 10 months in terms of agentic usage," said Dan McNamara, AMD's senior vice president and general manager of compute and enterprise AI. McNamara added that CPU sales are going to "really ramp."

GPUs think, CPUs do

Muse, Dots and even coding software like OpenAI's Codex use AI models to plan tasks that can run autonomously. The agents need their own computer that can run for hours or even days in the background, explained Futurum Group CEO Daniel Newman.

GPUs are still needed for inference, or processing an AI model, but an agent also needs a CPU.

"CPUs are actually performing the workflows while GPUs are doing the thinking," Newman said.

Meta and OpenAI use virtual machines, which partition a server into several lighter virtual computers, so one server can support scores of users.

Some AMD EPYC processors have up to 192 CPU cores. According to benchmarks and the agents themselves, Muse's virtual computer uses two cores and Dots uses nine.

Benchmarks and queries show Dots using an AMD EPYC 9V74 CPU, which is available from resellers for under $3,000. The EPYC 9D25 reportedly used by Meta for Muse costs even less on the secondary market. Nvidia GPUs can cost more than 10 times that for a single chip, and they're usually sold in clusters of hundreds or thousands of chips.

Investors are watching how Meta plans to serve Muse users, after downloads surpassed 5 million since last month's launch, according to Sensor Tower. Morgan Stanley estimates Muse's serving costs at $3 to $130 per month, with an average of $37 per user, depending on inference usage.

The firm says Meta's agent could account for 20% of AMD's 2026 chip sales.

There's a lot of money to be made in CPUs, and forecasts for the segment's growth are accelerating.

Futurum estimates $118 billion in total CPU sales in 2027, almost double the firm's previous forecast in May.

In July, AMD predicted the entire CPU market would hit $220 billion in sales in 2030, up from a previous 2025 forecast of $60 billion. AMD said in July that it expects to take more than half of the market.

AMD currently commands about 46% of the market for x86 CPU units, Mercury Research said in August. Intel is also seeing a surge in demand.

But AMD has a foothold with hyperscalers, which is where many of the virtual machines for agents will be deployed. It also has the benefit of using an x86 architecture for its chips, allowing agents to use older software without worrying about compatibility issues.

"AMD is probably getting the most of this because they have the largest market share and the cloud hyperscaler world for CPUs," Klein said, adding that he "wouldn't rule Nvidia out."

"If you are a company that can supply both types of chips, you're in the best position," Klein said.

WATCH: Helios is AMD's first AI system to rival Nvidia Vera Rubin — we got an exclusive, first look

Got a confidential news tip? We want to hear from you.

Get this delivered to your inbox, and more info about our products and services.