Anthropic's Compute Scramble Shows Why AI Capacity Is Now a Vendor Risk

In the space of three days this week, Anthropic was reported to have signed a $10 billion cloud deal with a startup founded this year, confirmed it is building its own chip design team, and showed up in two earnings reports as a major buyer of other companies' computing power. Taken together, the news shows how far the leading AI labs are now going to secure capacity, and why enterprises that depend on their models should treat compute supply as part of vendor risk, not a background detail.

What did Anthropic do this week?

On August 4, TechCrunch reported, citing a Bloomberg report based on unnamed sources, that Anthropic had agreed to a deal reportedly worth $10 billion with Volta, an AI cloud startup founded earlier this year. The agreement would provide computing capacity over six years from a 133-megawatt facility in Norway running Nvidia's Vera Rubin systems, with crypto-mining company Bitdeer helping to develop the site. Anthropic had not commented to TechCrunch at the time of publication.

The next day, Anthropic confirmed to TechCrunch that it is building a team to design custom AI chips, after Business Insider spotted a job listing for a "custom silicon team." The company said it plans to co-design hardware and models to make its systems faster and more efficient. TechCrunch noted that The Information had reported in July that Anthropic was considering Samsung as a manufacturing partner.

Anthropic also appeared in two suppliers' results. AMD's second-quarter earnings release on August 4 highlighted a strategic partnership to deploy up to 2 gigawatts of AMD's MI450 Series GPUs for Anthropic. And SpaceX, reporting its first quarter as a public company the same day, said its revenue nearly doubled on the back of compute deals with Anthropic and Google, along with Starlink growth, according to TechCrunch.

How big is the AI compute buildout behind these deals?

SpaceX's numbers give a sense of scale. Its AI segment brought in $2.56 billion in the second quarter, up from $737 million a year earlier, and the company said it had signed cloud services agreements totaling $14.1 billion in contracted sales, according to its results release. AI segment capital spending was $15.8 billion in the quarter alone, and nameplate compute capacity reached 1.4 gigawatts. The release did not name customers, but TechCrunch reported that the capacity came from data centers in and near Memphis originally built to train xAI's models.

These deals sit on top of commitments Anthropic described earlier in the year. In May, when it announced the SpaceX agreement, Anthropic listed compute arrangements with Amazon of up to 5 gigawatts, a 5-gigawatt agreement with Google and Broadcom, $30 billion of Azure capacity through a partnership with Microsoft and Nvidia, and a $50 billion investment with Fluidstack in US AI infrastructure. The same announcement doubled some Claude Code usage limits, a direct example of new capacity turning into more access for customers.

Why are AI labs building their own chips?

Anthropic is following a path others have taken. TechCrunch noted that OpenAI unveiled a custom inference chip built with Broadcom in June, that Google DeepMind has long relied on Alphabet's TPUs, and that Meta has been developing its own MTIA accelerators.

The logic is cost and control. Inference, the work of running models for users, grows with every new enterprise customer and every agent left running. A chip designed around a lab's own models can, in principle, serve each request with less power and at lower cost than general-purpose hardware. Owning the design also reduces dependence on any one supplier. The tradeoff is time: custom silicon takes years to design and manufacture, so the deals with Volta, AMD, SpaceX and the hyperscalers are what will carry demand in the meantime.

What does a lab's compute supply mean for enterprise customers?

For most enterprises, compute supply shows up in three places: rate limits, reliability and price. When a model provider is short on capacity, customers see throttling, slower responses at peak times and tighter usage caps. When new capacity arrives, limits rise, as Anthropic's May announcement showed.

Spreading supply across many vendors, chip architectures and countries has benefits for customers. It reduces the chance that one outage or shortage takes a service down. But it also raises questions enterprise buyers should ask:

• Where will my workloads run? A facility in Norway or a former training cluster in Tennessee may matter for data residency, contractual commitments and regulatory filings.
• Is performance consistent across hardware? Latency and throughput can vary depending on where a request is served. Enterprises with strict service-level or evaluation requirements should test.
• What happens if a supplier stumbles? Several of the partners in these deals are young companies. Ask how capacity would be replaced.

AI Connect speakers have long urged buyers to plan for that kind of change. During a AI Connect panel on future-proofing AI infrastructure, Stefan Leichenauer, VP of Engineering at SandboxAQ, warned that a chosen vendor "may not even exist two years from now," and urged teams to use open frameworks and test several providers. The advice applies equally to the infrastructure layer beneath model providers.

Should enterprises spread their bets across model providers?

The same logic that drives labs to diversify suppliers applies to their customers. Enterprises relying on one model family for critical workflows inherit that vendor's capacity constraints. Building applications so that models can be swapped, and routing simpler tasks to smaller or cheaper models, reduces exposure.

Mistral AI's Aaron Delp made the case for model choice during a AI Connect webinar on right-sizing enterprise agents, arguing that the best return comes from "the right model at the right time in the right place." Teams that can move workloads between providers also have more leverage in contract talks, a point echoed in AI Connect sessions on AI factories and private infrastructure for regulated workloads.

What to watch next

Three signals will show whether this buildout is paying off for customers. First, whether usage limits and latency for enterprise plans keep improving as new capacity comes online. Second, how quickly SpaceX's contracted cloud revenue converts, since its finance chief said on the earnings call that $6.7 billion of it ramps over six months beginning in October, according to TechCrunch. Third, whether Anthropic's chip effort produces a timeline or a manufacturing partner. For enterprises, the takeaway this week is simple: the reliability of the AI services they buy now depends on a supply chain that runs from chip foundries to power grids, and it is worth asking vendors how theirs holds up.

Frequently asked questions

What is Anthropic's deal with Volta?

TechCrunch, citing Bloomberg, reported a deal worth about $10 billion for six years of compute from a 133-megawatt Norway facility using Nvidia Vera Rubin systems. Anthropic had not commented.

Is Anthropic making its own AI chips?

Anthropic confirmed it is building a team to design custom chips so it can co-design hardware and models. No timeline has been announced.

Why does AI lab compute supply matter to enterprises?

Capacity constraints show up as usage limits, slower responses and pricing, so a vendor's compute supply affects service reliability for its customers.

Disclosure: AI Enterprise Magazine uses AI tools in its editorial process, including Anthropic's Claude models. Anthropic had no role in this article.

Filed under: Data & AI Infrastructure, AI Strategy & ROI, AI Governance & Trust