Broadcom loan to Anthropic capped at $42bn as US datacenter plans outrun chip supply

Broadcom has agreed to lend Anthropic up to $42bn to help pay for its AI infrastructure. That is according to Reuters, which on Thursday cited the AI company's IPO prospectus. The loan could cover about a third of the $125.2bn Anthropic has committed to a five-year lease of tensor processing units, the chips Google designs with Broadcom. Reuters said Broadcom and Anthropic both declined to comment.

The money would come through convertible notes that could convert into Anthropic shares. Broadcom can also name a financing partner, and Anthropic said it does not expect any notes to be sold before its IPO.

The filing also flags a structural problem: Broadcom is both Anthropic's hardware supplier and its lender, which the prospectus calls "potential conflicts of interest". Those conflicts, it warns, could affect Anthropic's access to the computing power it needs, since Broadcom's decisions on pricing and hardware could limit how much infrastructure the company can buy. Some payment or performance defaults could also make much of the lease obligations due at once, in which case Anthropic said it could make only limited use of the $42bn facility to cover them.

Anthropic put cash into a restricted account for Broadcom's benefit in April 2026, according to the filing, and may have to add more in certain circumstances. That same month it announced an expanded deal with Google and Broadcom for multiple gigawatts of next-generation TPU capacity starting in 2027. In August, Bloomberg reported Broadcom was seeking more than $60bn in debt to fund chips for Anthropic.

Reuters reports that Anthropic is on course to become Broadcom's largest compute customer in 2027. Broadcom projects AI chip revenue of about $115bn in fiscal 2027 and $230bn in fiscal 2028. Anthropic lost $42bn in 2025 as revenue grew 12-fold, its prospectus shows, and its IPO could value it at $2tn.

The arrangement mirrors Nvidia, which has used its balance sheet to support sales of its chips. This week the Financial Times reported Nvidia had talked to insurers about loans backed by its chips. Seaport Research analyst Jay Goldberg told Reuters that Nvidia "is putting in place a massive amount of its balance sheet, and Broadcom is having to follow suit". Robert Leitao, managing partner of Rothschild & Co, told Reuters it "feels that there's quite a concentrated bet right now on two companies being able to generate enough revenues to support all the financing that's happened".

The financing news lands on top of a supply problem that predates it. On 30 September, The Register reported on a Jefferies analysis, citing satellite analytics firm SynMax, which tracks land clearing and construction progress at datacenter sites week by week. SynMax estimates US deployments will rise from roughly 11 GW of gross capacity in 2025 to 16-18 GW in 2026, above its previous forecast of 14-16 GW. It puts the practical upper limit for 2027 in the low 20s of gigawatts.

That is well short of the more than 80 GW implied by some announced project pipelines and chip demand models for 2028. Land clearing for new projects has plateaued, and an earlier Jefferies report found only half the US capacity scheduled for 2026 was under construction, with work yet to begin on as much as 80 percent of the 2028 pipeline.

Advanced packaging is the second constraint. Fabricating accelerator dies is only part of the job: they must be packaged with components such as high-bandwidth memory before OEMs can install them in servers. Even chips manufactured on US soil may need to be sent overseas, typically to Taiwan, for packaging.

SynMax estimates existing advanced packaging capacity could support accelerators drawing roughly 13 GW of gross power. After accounting for CPUs, memory, cooling and other loads, that converts to about 17.5 GW of total datacenter power. Two additional packaging projects expected online in 2027 could support another 6 GW, taking the practical ceiling into the low 20s unless capacity expands faster than forecast.

China's two largest chip buyers are moving in the opposite direction. Huawei's rotating chairman Eric Xu claimed in a Q&A-style release this week that the company's Ascend line of neural processing units had exceeded Nvidia in Chinese market share. "It's pretty hard to collect data about the market share of Nvidia in China, but based on the data we have collected, Ascend has surpassed Nvidia," he said. Xu added that while Huawei's chips may be less advanced, "at least their supply is assured, so that you don't have to worry about chip supply day in and day out".

On Wednesday, DeepSeek open-sourced six software modules tailored for Huawei's Ascend chips, including an Ascend-compatible version of TileLang, its programming language for high-performance kernels. The SCMP reported the project's GitHub page now lists support for Ascend 950 accelerators with native code generation.

Nvidia's own numbers show how far the China business has fallen. During the company's Q2 earnings call, executives said shipments of H200 accelerators to the region amounted to less than 1 percent of datacenter revenues. Ars Technica reported on 28 September that China is now weighing allowing some of its biggest AI firms to import perhaps millions more Nvidia chips over the next year, with the Ministry of Industry and Information Technology asking Alibaba and ByteDance to detail plans to buy RTX Pro 5500 parts.

The design side of the market is consolidating around a different thesis. OpenAI and Synopsys signed a multi-year partnership to build GPT-Synopsys, a model intended to reason about chip design and verification and operate Synopsys' EDA tools directly. Engineers will delegate design objectives and review the output, the two companies said, with early tests already underway at semiconductor customers.

Tom's Hardware reported on 30 September that Cadence, Synopsys and Siemens EDA all now pitch agentic AI for chip design, largely built on Nvidia's stack, with autonomy claims that range from Level 4 to Level 5. The outlet noted that all the claimed speed improvements are the vendors' or their customers' own figures, with many carrying "up to" or "early evaluation" caveats, and that the only evaluator Synopsys named in July, AMD, has not provided analysis beyond an endorsement.

Elsewhere, the packaging bottleneck has a supplier response. Emergence AI told EE Times it is moving its neuroformal AI technology into active deployments with fabless semiconductor companies, with integrated device manufacturers asking it to extend the work into the fab. Co-founder Satya Nitta said the goal is more die per wafer rather than faster chips: "You cannot make chips any faster, but what you can do is definitely get more chips per wafer yielding with AI."

Not every chip announcement this week concerns accelerators. MaxLinear announced Puma 9, a DOCSIS platform it says is the first to support DOCSIS 3.1, 3.1+ and 4.0 on a single SoC, built on 8nm process technology and ARM rather than the Intel x86 architecture used in prior Puma generations. The company claims it cuts costs by 30 percent to 50 percent against its predecessor, adds DDR5 memory support to sidestep DDR4 supply pressure, and includes an integrated NPU for edge AI. Modems based on it are expected in 2027.

The common thread across the week's news is that money is no longer the binding constraint. Broadcom can commit $42bn to one customer, and Anthropic can commit $125.2bn to a single lease. What neither can do is conjure packaged accelerators faster than the fabs and packaging houses can produce them, or build datacenters faster than power and permits allow. The AI build-out is being priced like a balance sheet exercise and delivered like an industrial one.