Inside Anthropic's IPO: The $900 Billion Data Center Gamble

Deep News
6小時前

As of 2045, Anthropic has committed more than $900 billion in infrastructure spending, a substantial portion of which will flow into a project called AnthropicDirect: chips it controls, deployed in data centers built to its own specifications.

The goal of this approach is to reduce reliance on expensive cloud computing from Google, Amazon, and especially Elon Musk's SpaceX, thereby driving down costs. SpaceX's revenue has also swelled dramatically because of Anthropic's massive purchases.

But that means investors taking part in this week's Anthropic IPO must bet that the company can honor these long-term payment commitments no matter how its own large-model needs shift. Anthropic CEO Dario Amodei said last Friday during a fireside chat with JPMorgan Chairman Jamie Dimon: "Compute is the single most core operating challenge the company faces." He acknowledged that budgeting carries enormous uncertainty given how quickly technology and revenue are evolving.

The five-year-old company says that after the IPO closes, its cash and investments will reach $213 billion. Even so, Anthropic's prospectus is full of risk warnings about infrastructure expansion. The prospectus states that simultaneously juggling multiple deals with developers, chip lessors, and power suppliers will "expose us to structural timing mismatches and liquidity risk"; payments must be made before compute comes online, which is changing the company's cash flow structure.

Amodei and the other founders acknowledged in their letter to investors that this upfront compute investment "looks almost reckless at first glance," but that they will strive to be "bold yet rational."

As usage surges and bills climb, Anthropic needs to build its own cloud compute more than ever. The company's third-quarter compute costs reached $9.9 billion, more than all its other operating costs combined. About half of that went to expensive leased compute, such as its recent purchase agreement with SpaceX, and data centers like the one in Louisiana are precisely meant to replace that leased capacity.

That data center, developed by Hut8 and called River Bend, will come online next year and can provide Anthropic with 245 megawatts of compute. Google and Broadcom are supporting the project's launch: TPU chips the two co-developed will be deployed in large numbers in these machine halls. Google is providing guarantees for leases on the data centers housing the TPUs, helping the project secure lower-cost debt financing. Hut8 issued $3.25 billion of bonds at a 6.2% coupon for the River Bend project. Meanwhile, Broadcom also took part in financing on the chip side.

In June, BigSky, a private credit vehicle led by Apollo and backstopped by Broadcom, agreed to purchase an initial batch of roughly 875,000 TPU chips (about 1 gigawatt of compute), to be deployed across four data centers including River Bend. Documents seen by the media show that Anthropic must pay about $7.2 billion in annual rent and can buy the chips after five years at a price of $95. Anthropic also agreed with Broadcom to lease another 3.5 gigawatts of TPU compute; excluding interest, the total chip lease rent reaches $161 billion.

According to media calculations, once fully operational, this roughly 4.5 gigawatts of compute will cost about $32 billion a year, equivalent to $7 billion per gigawatt. By comparison, EvercoreISI estimates that, based on the pricing of Anthropic's first agreement with SpaceX, procuring 1 gigawatt of AI cloud compute from SpaceX (including machine halls and power) costs about $30 billion a year. An investor familiar with the investment banking models said that Goldman Sachs, the lead underwriter of this IPO, expects Anthropic to begin cutting back the scale of its SpaceX compute purchases starting in the second half of next year.

The potentially enormous cost savings are exactly why Anthropic is pushing so hard to bring facilities like River Bend online as quickly as possible. To move fast, Anthropic chose to work with former cryptocurrency miners such as Hut8, TeraWulf, and Cipher; these companies have spent years hunting for cheap power and grid access, and they are good at building machine halls quickly. A foreman who came to the River Bend site from North Carolina said: "Work never stops here; they are pushing desperately to move forward."

Many machine halls are sited in rural areas such as upstate New York, Texas, Indiana, and Kentucky, where power prices are lower and supply resources are easier to obtain. River Bend is just two miles from a nuclear power plant, and the site has high-voltage transmission lines and four natural gas pipelines. A year ago, this land was still used for logging and hunting. In February, construction crews began leveling this 280-acre hilly plot. The substation used to step down grid power and the steel frame of a 600,000-square-foot building were essentially completed in only about four months. Now, a crane workers call "Big Red" is hoisting 60,000-pound concrete wall panels.

Hut8 is expected to deliver the first data hall next spring; by then, Fluidstack, which represents Anthropic in leasing the site, will begin paying rent. Hut8 CEO Asher Genoot said the experience of building crypto mines showed them which parts of traditional data centers are "over-engineered" and can be simplified. The company also asks a lot of its employees; referring to one of the projects, he said: "Our workers haven't been home in seven months."

Heavy use of TPUs makes Anthropic unusual among AI labs — it has already moved far away from a single-track reliance on Nvidia. The prospectus notes that in-house software lets it train and run models on Google, Nvidia, AMD, and Amazon chips, choosing the hardware with the best cost-performance for different tasks. Anthropic co-founder Tom Brown said in an IPO video: "Early on, we made a very unconventional choice: a multi-chip path. We can match the most suitable chip to each type of task." He said that in the past, adapting to a new chip required dozens of engineers and months of work; this year, Anthropic's Mythos model deployed Claude on AMD processors in just three days.

The paid-customer business is already highly profitable, with gross margins above 85% in the nine months through September. But only about a quarter of Anthropic's compute supply serves paying customers. Most of the rest is used for training new models and research and development, and building its own data centers is aimed precisely at lowering that portion of costs.

The risk is that large-scale infrastructure brings huge fixed expenditures, equivalent to betting that large-model demand will keep growing rapidly enough to cover the bill. Such projects are also facing growing opposition. During this month's U.S. midterm elections, many voters backed candidates and ballot measures calling for halting data center construction, and market concerns about the power and water consumption of compute infrastructure continue to rise. For Anthropic, once construction is delayed, projects built on complex financing structures become extremely costly.

But at least in Louisiana, local attitudes toward Anthropic's big bet are supportive. Louisiana Governor Jeff Landry said last month at the River Bend project ceremony: "One of the most admirable things about Americans is that they do not fear challenges. America is the most risk-taking country in the world today."

免責聲明:投資有風險,本文並非投資建議,以上內容不應被視為任何金融產品的購買或出售要約、建議或邀請,作者或其他用戶的任何相關討論、評論或帖子也不應被視為此類內容。本文僅供一般參考,不考慮您的個人投資目標、財務狀況或需求。TTM對信息的準確性和完整性不承擔任何責任或保證,投資者應自行研究並在投資前尋求專業建議。

熱議股票

  1. 1
     
     
     
     
  2. 2
     
     
     
     
  3. 3
     
     
     
     
  4. 4
     
     
     
     
  5. 5
     
     
     
     
  6. 6
     
     
     
     
  7. 7
     
     
     
     
  8. 8
     
     
     
     
  9. 9
     
     
     
     
  10. 10