Anthropic's $518B AI Buildout: Who Wins and Loses Among Amazon, Google, Microsoft, AMD, and Broadcom?

 

A prospectus that has never been published has given the market its first clear look at the size of the bill the industry's most closely watched AI lab has signed. According to Reuters, which reviewed the confidential IPO filing, Anthropic expects to spend at least $518 billion over roughly a decade building AI infrastructure with six partners, and about 80% of that amount is either non-cancelable or payable regardless of usage.

 

The figure matters to the whole supply chain because it is the first time the compute commitments scattered across separate press releases have been pulled into one traceable ledger. Google, Amazon, Microsoft, Broadcom, AMD and SpaceX unit xAI all appear in the same document, variously as cloud providers, custom silicon suppliers, equipment lessors and investors. For anyone holding those stocks, the question is no longer whether Anthropic will spend. It is on what schedule, under what legal form, and on whose revenue line.

 

One caveat belongs up front. The prospectus remains confidential and Anthropic has not published these figures. Every financial and contractual detail below comes from reporting by Reuters and other financial outlets on a non-public document, and none of it has been confirmed by the issuer or by regulators.

 

 

 

$518 billion is not a single payment. It is the aggregate of long-term contractual obligations spread across six partners over roughly ten years, averaging somewhere in the region of $50 billion a year rather than landing in any one period.

 

Roughly 80% is locked. Reuters reports that this portion cannot be canceled, or must be paid whether or not the capacity is used. The $31.4 billion Microsoft commitment can be terminated only in the event of Microsoft's uncured material breach.

 

Broadcom carries the largest single exposure. About $161.2 billion of equipment lease obligations, tied to custom silicon and networking hardware, and largely non-cancelable.

 

The xAI terms are the outlier. Agreements worth up to $84.5 billion for Nvidia-based capacity through 2029 are largely cancelable on 90 days' notice, a structure nothing else on the list matches.

 

The binding constraint is shifting toward memory and power. With HBM capacity sold out and conventional DRAM pricing sharply higher, memory suppliers are the overlooked link in this chain.

 

 

 

Reuters reports that Anthropic submitted a draft S-1 confidentially to the U.S. Securities and Exchange Commission in June. SiliconANGLE's summary of the document puts 2025 revenue at nearly $4.6 billion, roughly twelve times the 2024 figure, against an operating loss of $8.06 billion, with computing infrastructure accounting for more than 91% of that loss. PYMNTS, compiling the Reuters reporting, adds a net loss of $42 billion, cash and short-term investments of $20.28 billion at year-end, and compute and infrastructure spending of $7.33 billion, close to triple the prior year.

 

The balance of the document is telling in its own right. Reuters notes that of the prospectus's 261 pages, roughly 80 cover risk factors against about 48 describing the business. The company tells investors that demand for advanced AI systems is likely to exceed available supply and that development will be limited principally by the availability of compute, which is the stated rationale for committing so much capital so far in advance.

 

 

Aggregating the individual figures reported by Reuters produces the structure below. The amounts come from media accounts of a confidential document, are disclosed on differing bases and over differing terms, and are not strictly comparable.

 

 

 

The point deserves emphasis because it determines how the valuation and the risk should be calculated. The $518 billion is the sum of multi-year contractual infrastructure commitments. It is not a single year's capital expenditure and it is not money already paid. TheStreet's reconstruction of the timelines shows the Google obligation running from April 2026 to July 2033, the Amazon obligation from May 2026 to April 2036, and the Microsoft obligation from November 2026 to May 2033. Spread across a decade, the total implies something on the order of $50 billion a year.

 

For suppliers, that means revenue recognition follows capacity delivery over years rather than clustering in a quarter. For Anthropic, it means a set of long-dated commitments and lease obligations sitting against a year-end 2025 cash position of $20.28 billion. The gap between those two numbers is precisely why the document has drawn the attention it has.

 

 

The reasoning set out in the filing is straightforward: compute availability is becoming the main constraint on model capability. The revenue trajectory supports it. In April's announcement of the expanded Google and Broadcom partnership, Anthropic said run-rate revenue had passed $30 billion, up from roughly $9 billion at the end of 2025. Reuters, citing the prospectus, reports the company projects revenue of roughly $190 billion to $200 billion for 2028.

 

That growth translates mechanically into capacity needs. Training a frontier model requires densely deployed accelerator clusters, while inference demand scales with usage and cannot be deferred the way a training run can. On the supply side, industry estimates put the cost of a one-gigawatt data center near $50 billion, of which roughly $35 billion is typically chips, a framing CNBC used in covering the Google TPU agreement. Reserving multiple gigawatts of that capacity effectively requires multi-year, non-cancelable contracts, because suppliers need that certainty before committing to fabs and procurement.

 

The price is flexibility. Non-cancelable terms transfer demand uncertainty onto Anthropic itself, and if revenue growth undershoots, the obligations do not shrink with it. That is the tension the risk section spends 80 pages on.

 

 

 

The Amazon relationship deepened considerably in April. According to TechCrunch's account of the agreement, Anthropic committed to spending more than $100 billion on AWS technologies over ten years, covering Trainium2 through Trainium4 plus Graviton CPUs, while Amazon invested a further $5 billion. Data Center Dynamics reports that Anthropic will lease up to 5 gigawatts of AWS capacity, and that Project Rainier, the joint cluster that went live in October 2025, was built around nearly 500,000 Trainium2 chips.

 

For Amazon the value extends past revenue. Anthropic is the reference customer proving that in-house silicon can carry frontier training workloads, which bears directly on AWS's pricing power in the accelerator market. On the financial side, The Motley Fool's analysis of the AWS backlog notes the backlog stood at roughly $496 billion as of June 30, with the Anthropic commitment averaging more than $10 billion a year. That is visibility and concentration in the same number.

 

 

Google carries the second-largest figure in the table at about $111.1 billion. The relationship dates to 2023 and expanded sharply in October 2025, when CNBC reported Anthropic had secured access to as many as one million TPUs in a deal worth tens of billions of dollars, bringing well over a gigawatt of capacity online in 2026. In April, Anthropic, Google and Broadcom signed for multiple additional gigawatts of next-generation TPU capacity. Broadcom's Form 8-K filed with the SEC describes roughly 3.5 gigawatts accessed through Broadcom beginning in 2027, and states plainly that consumption of the expanded capacity depends on Anthropic's continued commercial success.

 

That sentence is worth reading twice. It writes demand risk into a supplier's public disclosure, and it explains why the market began reassessing the quality of TPU-linked revenue as soon as the prospectus figures surfaced.

 

 

Microsoft's $31.4 billion is the smallest line in the table and the tightest in its terms. Reuters reports the commitment runs from November 2026 to May 2033 and is non-cancelable except in the event of Microsoft's uncured material breach. It traces back to the November 2025 arrangement in which, per the joint announcement from Anthropic, Microsoft and Nvidia, Anthropic committed to purchase $30 billion of Azure compute capacity and to contract up to one additional gigawatt, while Nvidia and Microsoft committed to invest up to $10 billion and up to $5 billion respectively.

 

The three hyperscalers simultaneously act as investors, customers, cloud providers, distributors and competitors, and the filing names that as a risk, with Reuters citing language that these incentives may not be fully aligned with Anthropic's interests. A related data point: TrendForce, citing Reuters, reports that 47% of Anthropic's sales last year were routed through Amazon and Google, up from 32% in 2024 and 11% in 2023.

 

 

The roughly $161.2 billion of equipment lease obligations makes Broadcom the single largest exposure on the list, and the legal form differs from the cloud contracts. These are leases against physical racks and custom accelerators, not metered cloud services.

 

Broadcom's own results already reflect the demand. Its third quarter fiscal 2026 results show revenue of $29.6 billion, up 86% year over year, with fourth-quarter guidance of approximately $34.8 billion. Yahoo Finance adds that AI semiconductor revenue rose 221% to $16.7 billion in the quarter, that management expects Anthropic to become its largest XPU customer in 2027, and that it projects AI semiconductor revenue of about $115 billion in fiscal 2027 and $230 billion in fiscal 2028.

 

The financing structure deserves attention too. MarketBeat describes the XPV platform Broadcom built with Apollo and Blackstone, under which a managed fund pays Broadcom for the chips as Anthropic deploys them and then receives lease payments from Anthropic over time. The arrangement front-loads the procurement capital, which also means that if the lessee struggles to pay, the racks have to find a secondary market. For Broadcom shareholders, the financial transparency that follows an Anthropic listing will directly change how verifiable this revenue stream is.

 

 

AMD's reported scale is modest relative to the others, but the structure is characteristic of this cycle. Per Bloomberg's report on the agreement, Anthropic will buy as much as 2 gigawatts of AMD's Instinct MI450 chips beginning in the first half of 2027, while AMD committed to invest up to $5 billion tied to deployment milestones. Reuters, citing the prospectus, puts the compute capacity AMD is expected to supply at more than $20 billion. A supplier selling chips while funding the buyer has become a recurring pattern, and it both accelerates deployment and complicates any assessment of revenue quality.

 

Nvidia is a different case. It carries no standalone contract figure in the table because its exposure flows through partners. The Azure capacity runs on Grace Blackwell and Vera Rubin systems, and the capacity available through xAI is likewise Nvidia-based. Each round of Anthropic diversification therefore expands Nvidia's indirect order book while simultaneously funding the alternatives, whether TPU, Trainium or Instinct. That two-way relationship makes any single order announcement a poor proxy for the whole picture.

 

The xAI agreement is also the loosest set of terms on the list. Reuters reports that commitments of up to $84.5 billion running through 2029 are largely cancelable on 90 days' notice. Placed next to Broadcom's lease obligations, it shows how unevenly Anthropic has preserved optionality across its suppliers.

 

 

Discussion of compute contracts usually stops at the accelerator, but through 2026 the binding physical constraint has been moving toward memory. S&P Global Market Intelligence research describes how Samsung, SK hynix and Micron have tilted capacity toward HBM, tightening conventional DRAM supply and pushing prices higher, with HBM's superior margins reinforcing the allocation.

 

Anthropic's response has been to put memory under contract as well. According to Micron's announcement, the June agreement spans memory and storage architecture co-design, a supply agreement across Micron's data center portfolio, enterprise adoption of Claude inside Micron, and a strategic investment in Anthropic's Series H round. Micron's own numbers show the intensity of the cycle: its third quarter fiscal 2026 results reported revenue of $41.46 billion against $9.30 billion in the same period a year earlier.

 

Korea has been equally active. Investing.com reported in July that Samsung Electronics and SK hynix announced roughly $950 billion of US supply partnerships, including SK hynix's planned $750 billion of memory supply to American companies such as Nvidia, and noted that both had discussed supply agreements with Anthropic. Separately, the memorandum of understanding between Samsung and Broadcom is expected to exceed $200 billion over five years through 2030, covering HBM supply, 2-nanometer-and-below foundry and advanced packaging for Broadcom's AI accelerators, with Anthropic among the most important end customers of that line.

 

This is why the $518 billion does not stop at three cloud providers. Custom silicon orders propagate upstream through packaging and memory, and because HBM capacity additions typically take twelve to eighteen months, the pace at which new memory supply arrives is often what actually sets the deployment schedule.

 

For investors who want to turn a supply chain view into a position, the stock trading section on MEXC covers most of the US names above, which keeps the cross-asset work inside a single account.

 

 

The legal form of a contract determines where the risk lands. Under normal conditions, non-cancelable terms favor the supplier by transferring demand uncertainty to the buyer and improving revenue visibility. But the transfer does not eliminate risk. It relocates it.

 

The first layer is Anthropic's own liquidity. Roughly 80% of the commitments are non-cancelable or take-or-pay, against $20.28 billion in cash and short-term investments at the end of 2025. The balance should be materially higher following the Series H round in May, but relative to a decade of contracted obligations, the gap still has to be closed through continued fundraising, IPO proceeds and revenue growth.

 

The second layer is physical delivery. TechTarget notes that contracts alone cannot guarantee that power and data center capacity arrive on schedule. Between contracted gigawatts and energized gigawatts sit siting, interconnection, transformer lead times and construction, and a delay anywhere in that chain opens a gap between contract value and recognized revenue.

 

The third layer is transmission. If a lessee cannot pay, Broadcom can resell the racks, but there is little historical precedent for the price or speed of a secondary market in used AI racks. By the same logic, a default against a cloud backlog would damage not only one customer's revenue line but confidence in how backlog is reported across the sector.

 

The fourth layer is a mismatch of horizons. The contracts run seven to ten years while model architectures, chip generations and inference cost structures are all moving quickly. Locking in payment obligations through 2033 against a 2027 technology roadmap is itself a bet on the stability of that roadmap.

 

 

Put together, a few directional conclusions follow.

 

Revenue visibility across the supply chain has genuinely improved. Where the market previously inferred demand from capex guidance and management commentary, there is now a document setting out amounts, terms and cancellation conditions, which is real information for assessing AI-linked revenue at Broadcom, AMD and the three hyperscalers.

 

Concentration risk has been exposed in the same motion. Broadcom has said publicly that its business is concentrated among six frontier model customers, with Anthropic expected to be its largest XPU customer. When downstream demand sits with a handful of companies that are not yet profitable, supplier revenue quality becomes tied to those customers' ability to raise capital. An Anthropic listing, with quarterly disclosure, makes that linkage verifiable, which is a neutral-to-positive development for the chain as a whole.

 

For secondary market participants, three data series carry more signal than headline contract values: the size and composition of cloud backlog additions, whether Broadcom and AMD revenue recognition tracks the known deployment schedule, and the direction of HBM and conventional DRAM contract pricing. Together they determine whether contracted dollars become cash flow. The adjacent question of how AI reshapes incumbent software and services markets runs on a separate track, and the analysis of why IBM shares moved on Anthropic's COBOL news is a useful case study there.

 

The listing date remains an open variable. TrendForce, citing Reuters sources, reports that Anthropic could begin marketing its IPO as early as mid-October with a listing potentially completed days before the November US midterms, though no date has been confirmed. For a read on how the market is pricing that timing, the walkthrough of Polymarket's Anthropic IPO odds offers another angle.

 

The market has already set its own price ahead of any real quote. CryptoSlate reports that Anthropic pre-IPO perpetual contracts traded $643 million through September 21, against $590 million for all of August, with roughly $80 million of open interest. These contracts settle in cash, convey no ownership, and could diverge sharply from an actual share price once one exists.

 

The bill is now on the table and the market is pricing this listing with real money. Trade the ANTHROPIC perpetual on MEXC and turn your read on the biggest compute bet in tech into a position you can actually size

 

 

What this document changes, in James Mitchell's reading, is not the market's view of AI demand but how the quality of that demand can be assessed. For two years the chain has run backwards, inferring end demand from supplier order announcements, a method that leaves ample room for the same dollar to be counted twice. A bottom-up purchase ledger now exists, and for the first time the two sides can be checked against each other. That is a shift from rumor-driven to document-driven analysis.

 

The likeliest misreading is treating $518 billion as either a valuation denominator or an immediate cash problem. It is neither a single year's spending nor money already paid, but a set of obligations with different terms and different horizons. What deserves unpacking is the terms themselves: the Broadcom leases are largely non-cancelable, the Microsoft contract is terminable essentially only on the counterparty's breach, and the $84.5 billion xAI arrangement can be exited on 90 days' notice. Cancelability, not headline value, determines whether a commitment behaves as a rigid liability or an adjustable expense under stress. Summing all three as though they were the same instrument is the easiest error available here.

 

The second underappreciated point is the mismatch between contract timing and physical constraints. Agreements are signed today for capacity arriving in 2027 and beyond, while HBM capacity additions typically take twelve to eighteen months and grid interconnection and data center construction take longer still. A lag between contract value and revenue recognition is therefore structural, which means a supplier missing a quarter does not necessarily signal softening demand, only a shift in the delivery schedule. The inverse also holds: recognition running visibly ahead of the known deployment timetable is worth questioning.

 

Three variables are worth tracking from here. The first is Anthropic's initial public disclosure after listing, particularly operating cash flow and the maturity profile of its commitments, which will let outside investors verify the reported figures for the first time. The second is the composition of cloud backlog additions, because backlog concentrating further into a handful of AI labs raises the credit exposure embedded in reported revenue. The third is the memory price curve, where HBM and conventional DRAM contract pricing serves as the closest thing to a real-time supply and demand reading for the chain.

 

Viewed across asset classes, the structure rhymes with earlier infrastructure cycles in telecoms, rail and data centers, where downstream buyers locked in capacity through long contracts, upstream suppliers expanded against those contracts, and the real test arrived when demand growth slowed while obligations continued to run. That does not make the same outcome inevitable here, since frontier AI revenue is still accelerating. It does suggest a standard for judgment: supplier visibility comes from contracts, and contract reliability ultimately comes from the buyer's ability to pay. Until Anthropic lists and begins reporting quarterly, that link in the chain remains the opaque one.

 

One note for investors who lean on models for research. Even the most capable language model cannot derive undisclosed contract terms from public information, and the discussion of what AI can and cannot do in price forecasting is worth reading before sizing anything.

 

 

 

No. The figure is the aggregate of contractual infrastructure obligations spread across six partners over roughly a decade, implying something near $50 billion a year rather than any single cash outflow. The terms differ by partner, with the Google obligation running to July 2033, the Amazon obligation to April 2036 and the Microsoft obligation to May 2033. Treating it as a lump sum badly overstates near-term pressure while understating how rigid the long-term commitment is.

 

 

No. The details come from Reuters and other financial media reporting on a confidential IPO prospectus. Anthropic has not published the figures and the prospectus itself remains non-public, having been submitted confidentially to the SEC in June. Until the document is publicly filed, every amount, term and condition described should be treated as a journalistic account of a private document rather than issuer-confirmed information.

 

 

By reported amount, Broadcom leads with roughly $161.2 billion of equipment lease obligations, ahead of Google at about $111.1 billion and Amazon at about $110 billion. Size alone is not the whole story, because the legal form differs. Broadcom's exposure is to physical rack leases while the cloud providers hold service contracts, and the two behave very differently in a default scenario.

 

 

Reuters reports that about 80% of the commitments either cannot be canceled or must be paid regardless of actual usage. The Microsoft contract is the clearest example: Anthropic cannot exit it except in the event of Microsoft's uncured material breach. By contrast, the xAI agreement worth up to $84.5 billion is largely cancelable on 90 days' notice. The same nominal dollar carries very different rigidity depending on those terms.

 

 

Because its exposure runs through partners rather than a direct contract with Anthropic. The Azure capacity is built on Grace Blackwell and Vera Rubin systems and the xAI capacity is Nvidia-based. At the same time, Anthropic is buying Google TPUs, AWS Trainium and AMD Instinct at scale, which are alternatives. Each expansion therefore adds to Nvidia's indirect order book while also funding competing architectures.

 

 

Because accelerators cannot deploy without HBM and storage alongside them. Samsung, SK hynix and Micron have shifted capacity toward HBM, tightening conventional DRAM supply and lifting prices. Micron has signed a memory and storage supply agreement with Anthropic and invested in its Series H round, and the Samsung and Broadcom memorandum of understanding feeds the same accelerator line. Since HBM capacity additions take twelve to eighteen months, memory is often the real gate on deployment timing.

 

 

Anthropic is not publicly traded, so there is no ticker and no share price. Indirect routes include holding listed companies that have invested in or supply the company, such as Amazon or Alphabet, or trading cash-settled pre-IPO derivatives. Those contracts represent no ownership whatsoever, their prices reflect market expectations rather than an actual quote, and a listing could reveal a substantial gap between the two.

 

 

Not yet. Reuters sources indicate marketing could begin as early as mid-October with a listing potentially days before the November midterms, though other reporting has suggested the debut may slip past that date. No exchange, ticker, price range or listing date has been announced, and any circulating timetable should be treated as unconfirmed.

 

 

The information above is provided for general market information and analysis only and does not constitute investment advice, financial advice, legal advice, tax advice or a recommendation to trade. The contract values, financial figures and timelines cited derive principally from media reporting on a confidential document that has not been publicly filed, have not been confirmed by the issuer or by regulators, and may differ from whatever is eventually disclosed. Prices of crypto assets, equities, pre-IPO derivatives and other related financial instruments can fluctuate sharply, and past performance, technical indicators and on-chain data do not guarantee future results. Pre-IPO perpetual contracts settle in cash, represent no equity ownership in any company, and may diverge materially from any future listed share price. Readers should conduct their own research and make decisions based on their own financial circumstances, investment objectives and risk tolerance, consulting a qualified professional where appropriate. The MEXC Crypto Pulse team accepts no liability for any direct or indirect loss arising from the use of this information.

 

 

James Mitchell specializes in technical analysis, market trends, and trading strategies for both Bitcoin and altcoins. Based in London, he has over 10 years of experience in financial markets. Before joining MEXC Learn, James worked as a senior analyst at a leading European investment firm, where he developed expertise in risk management and quantitative trading. His transition to cryptocurrency markets began in 2017, and he has since become recognized for his data-driven approach. He holds a Master's degree in Financial Economics from the London School of Economics. His analytical approach combines traditional technical analysis with on-chain metrics to provide readers with actionable insights.

 

Areas of Expertise: Technical Analysis, Market Trends & Cycles, Trading Strategies, Bitcoin & Altcoin Analysis, Risk Management.

 

 

 

Want the fastest access to MEXC's latest updates? Join our official Telegram group now!

MEXC Platform: Download App | Zero Fee Trading | Buy USDT Now

Join MEXC Community: X (Twitter) | Telegram | Discord

Get Insights: MEXC Learn | Terms of Service | MEXC Blog | MEXC News

Account Verification: Understand KYC | How to Complete KYC

External Content Platforms: Substack | Medium | Paragraph | LinkedIn | X(News)