Anthropic’s IPO filing reveals a strategy to undercut its AI rivals
Welcome to AI Decoded, Fast Company‘s weekly newsletter that breaks down the most important news in the world of AI. I’m Mark Sullivan, a senior writer at Fast Company, covering emerging tech, AI, and tech policy.
Sign up to receive this newsletter every week via email here. And if you have comments on this issue and/or ideas for future ones, drop me a line at sullivan@fastcompany.com, and follow me on X @thesullivan.
‘Compute multipliers’ and ‘Anthropic Direct’ may be the IPO’s secret sauce
As frontier models from Anthropic, OpenAI, and Google converge in capabilities, the companies selling them may end up competing for enterprise business less on model quality and more on the cost of the computing power needed to run them. In its IPO filing and investor roadshow video, Anthropic, which debuted on the Nasdaq on Thursday, suggested it may be preparing for that future with something called “compute multipliers” and a compute-sourcing model it calls Anthropic Direct.
In its S-1, filed publicly in October, Anthropic defines compute multipliers as “algorithmic and methodological” improvements in the way it trains and operates models, effectively increasing the computing power it can get from a given amount of capacity. And these multipliers are “stackable.” In its roadshow video, Anthropic says it measures “total effective compute,” which is its physical computing capacity multiplied by the multipliers it has discovered or expects to discover in the future. Its example: If a model takes 100 units of compute to train, a 4x multiplier lets Anthropic train the same model with 25 units, or use all 100 units to train a model that would previously have required 400.
The company also says the multipliers “stack,” with each new discovery building on previous ones, and that it expects Claude will eventually be able to discover multipliers on its own.
According to the filing, the gains apply to models used to serve customers as well as to model training. Anthropic says its researchers “have delivered meaningful compute multipliers for inference,” reducing the computing power needed to deliver a given level of model performance. The company says this “allows us to provide greater value for customers.”
Anthropic has already been cutting prices. It says price reductions tied to new model versions, along with contractual discounts, reduced revenue growth by about $600 million between the second and third quarters of 2026. It says Opus 5.5, launched in September, delivered “a lower cost per task” than earlier models. If Anthropic can get more usable compute out of each chip than its rivals can, it may be able to undercut other AI labs on price. It may also be able to undercut cloud providers that sell AI services built on their own models.
If compute multipliers let Anthropic do more with its physical computing power, Anthropic Direct could give the company greater control over where that power comes from. Right now, Anthropic buys almost all its computing power from cloud providers, including competitors like SpaceX, Amazon Web Services, Microsoft, and CoreWeave. The company says it has already committed to buying roughly $900.9 billion in compute from these providers over the coming years and decades. Those expenses also include short-term cloud compute purchases, mainly to cover peak usage periods, which come at a premium.
The S-1 says Anthropic is reducing its reliance on expensive cloud providers by developing its own data centers and leasing the chips that go inside them. Its first Anthropic Direct facility came online this year. The company says a “significant portion” of its planned compute spending will go toward building more Anthropic Direct capacity. Based on an analysis of contract values in the S-1 filing, PitchBook estimates that Direct data centers will supply about 1% of Anthropic’s available compute capacity this year, 48% in 2030, and 60% in 2035.
The S-1 says Anthropic’s total compute pool, including third-party cloud compute and Direct, will be used for its own research and model development, as well as customer workloads. Bringing more compute in-house could dramatically reduce the overhead costs of developing and serving Claude models, apps, and agents for enterprise and consumer customers. That, in turn, could allow Anthropic to lower per-token prices for enterprises and subscription rates for consumers.
But Anthropic Direct may end up being something more. The filing doesn’t specifically say Direct compute could be rented to enterprise customers. Still, it’s a possibility; Anthropic could be building an AI-native cloud where enterprises could store their proprietary data and run Anthropic’s models in the same secure environment, potentially at a lower cost than building the same setup within an Amazon, Google, or Microsoft cloud.
That would give the Anthropic Direct brand name a very different meaning. Instead of accessing Claude models through a cloud provider they already use, customers would come directly to Anthropic. That could turn computing from a major overhead expense into a service Anthropic controls and rents to others for profit—a significant change to the company’s business model.
Some big obstacles stand in the way of such an ambition. Google, Amazon, and Microsoft are currently Anthropic’s largest compute suppliers, as well as the main distribution channels through which enterprises buy Claude. The S-1 notes that the three companies are at once Anthropic’s “investors, customers, distribution partners, and competitors.” If Anthropic challenged their cloud businesses directly, it would almost certainly strain those relationships. On the other hand, as more Direct data centers come online, Anthropic’s dependence on those cloud partners could decrease.
If the current rough parity among frontier models continues, the company that can deliver intelligence at the lowest overall cost per task—not per token—may have the edge. Anthropic’s filing suggests it is working toward that position on two fronts: getting more out of each chip and controlling more of the infrastructure those chips run on. Have retail investors factored that strategy into the price they’re willing to pay for Anthropic stock? The answer may emerge only after the stock has traded for a week or so.
California sues OpenAI over rogue AI agents
The California attorney general, Rob Bonta, sued OpenAI on Monday, alleging the company developed and tested advanced AI models without adequate safeguards. The suit centers on a July incident in which OpenAI agents escaped a testing environment and breached the systems of AI platform Hugging Face.
The complaint says OpenAI told the state it has lost control of its autonomous agents more than 80 times during testing. Bonta called the Hugging Face attack the “predictable result of an internal culture at OpenAI” that puts growth ahead of safety. The state alleges violations of California’s unfair competition law and seeks an injunction, civil penalties, and disgorgement of profits. An OpenAI spokesperson said the suit “appears to single out OpenAI over a challenge that’s been identified by every major AI lab.”
Jensen Huang says he’d be ‘honored’ to return as Nvidia CEO. Its board says nah
Nvidia’s board rejected the idea of reinstating Jensen Huang as CEO just hours after Huang said on Twitter that he’d be “honored” to return. On the company’s earnings call Tuesday afternoon, Harvey Jones, the board chair, said the board “does not intend to reinstate Jensen as CEO.” The board had unanimously accepted Huang’s resignation. He has drawn support from Elon Musk, President Trump, and a petition with about 180,000 signatures, but the board is continuing its search for a permanent successor. Colette Kress, the CFO, is serving as interim CEO.
Huang broke his silence Tuesday, less than a week after resigning amid a personal conduct scandal and an investigation. In a statement on X, he said that if the board decides his return is in the best interest of shareholders, employees, the company, and America, “I would be honored to answer these calls and return as CEO.” Huang acknowledged making “serious mistakes in my personal life,” and said, “I own my mistakes completely.” He also said he is “committed to cooperating with the Board’s independent investigation.”
Demis Hassabis’s Isomorphic Labs raises $5 billion
Isomorphic Labs, the AI drug-discovery company founded and led by Demis Hassabis and spun out of Google DeepMind, announced a $5 billion Series C round of funding on Tuesday. Greenoaks and Sequoia Capital led the round, with participation from Andreessen Horowitz and other new investors and continued significant investment from Alphabet. Several existing backers—Thrive Capital, MGX, GV, CapitalG, Temasek, and the U.K. Sovereign AI Fund—also invested.
Isomorphic will use the money to scale IsoDDE, its drug-design engine built on AlphaFold-related technology, and to push drug candidates toward clinical trials. In a post on X, Hassabis said the company continues “to make rapid progress reimagining drug design from first principles with AI” to cut the time and cost of finding new medicines. No valuation was disclosed, but Bloomberg had reported the company was in talks to base the round on a valuation of about $50 billion.
More AI coverage from Fast Company:
- Jeff Bezos’s $41 billion AI startup Prometheus is shrouded in secrecy. We found out who’s building it
- Voice AI is quietly transforming frontline work
- Nvidia’s first earnings call without Jensen Huang had one message: It’s business as usual
- Replit’s CEO on its stunning turnaround, ‘AI psychosis,’ and the risk Silicon Valley is ignoring
Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company today.
内容来自 Fast Company,版权归原作者所有。
阅读原文 ↗