May 6, San Francisco, Anthropic Developer Conference. Chief Product Officer Ami Vora did not release a new model on stage. She said that within a month, the company will take over all the computing power of SpaceX's Colossus 1 data center, exceeding 300 megawatts and 220,000 NVIDIA GPUs.
The audience fell silent for a few seconds. The developers were waiting for model updates, but Anthropic delivered computing power.
Six days later, The New York Times reported that Anthropic was negotiating a new round of financing with a valuation of up to $950 billion. Bloomberg later added that the financing size was between $30 billion and $50 billion, potentially completed before the end of May, but the term sheet had not yet been signed. If successful, Anthropic would surpass OpenAI's record $852 billion valuation set in March this year.
It took only three months from the $380 billion valuation in the Series G round in February to $950 billion in May. This valuation curve has no precedent in the history of technology business.
But what is truly worth questioning is the transfer of power happening behind these numbers: cloud providers and computing power giants are handing over their most scarce resources to the same company. An AI lab founded just five years ago has, in an extremely short time, become the de facto controller of the entire AI infrastructure layer.
The title 'Suzerain State' is apt.
01
Four Tributes, One Suzerain State
On April 21, Amazon announced an additional investment of up to $25 billion in Anthropic. In return, Anthropic committed to spending over $100 billion on AWS in the next decade, covering the Trainium series of chips.
Amazon disclosed at the same time that its 2025 investment of $8 billion in Anthropic was now valued at over $70 billion. Of course, at least in accounting terms, this gain was treated as unrelated to the commercial relationship.
Three days later, on April 24, Google followed suit, announcing an immediate cash injection of $10 billion and committing up to an additional $30 billion upon Anthropic reaching performance milestones, with a total ceiling of $40 billion. Simultaneously, Google Cloud pledged to provide about 5 gigawatts of computing power over the next five years, while Anthropic committed to spending $200 billion on Google Cloud during that period.
In May, SpaceX joined. The full computing power of Colossus 1 would be connected to Claude within a month.
Combined with the Azure capacity previously provided by Microsoft and NVIDIA, and the custom TPU chip cooperation involving Broadcom, the total computing power commitment secured by Anthropic in less than six months exceeded 20 gigawatts.
But the bill is not one-way. Amazon's and Google's money "every cent comes with a kickback clause." The financing Anthropic receives must be spent back on the investors' cloud services and chips at a scale of hundreds of billions of dollars. This is not the VC-style "giving you money to burn"; it is closer to compute suppliers finding a major customer for their own production capacity, using financing to lock in demand, and using demand to absorb capacity. Beneath the venture capital facade lies an infrastructure pre-sale contract.
A late April report by Fortune magazine exposed another layer of awkwardness: nearly half of the "impressive AI profits" for Google and Amazon in the first quarter came from the paper appreciation of their holdings in Anthropic, not from their own operations. For Amazon, this investment gain was 1.4 times the profit of its AWS. The quarterly earnings of the world's two largest cloud providers are being led by the valuation of a single startup.
At this point, the support structure of the Suzerain State fully emerges: SpaceX provides immediate GPU computing power, alleviating Claude's peak-time throttling; Amazon and Google provide capital, in-house chips, and future computing power, while distributing the Claude model through their cloud platforms and taking channel commissions; Microsoft provides Azure capacity, and although deeply tied to OpenAI, Claude is already one of the main models on its Foundry platform; Broadcom provides the hardware path of custom TPU chips, preventing Anthropic from being completely dependent on NVIDIA.
Four streams of tribute, one Suzerain State.
Each party believes it is using Anthropic to achieve its strategic goals: Amazon wants to fill Trainium chip capacity with Claude, Google wants to use Anthropic to contain the Microsoft-OpenAI alliance, Microsoft wants to prevent client loss to AWS due to the inability to use Claude, and SpaceX wants rental returns for the idle Colossus 1.
But each party, while delivering resources, is also handing over bargaining power. When all giants stake their strategic assets on the same company, that company is no longer an optional partner—it becomes a necessity that none can afford to lose.
02
Clashing Blades on the Eve of the IPO
The reason the scale of support continues to escalate lies in the fact that Claude's growth rate has exceeded everyone's expectations.
On April 7, 2026, Anthropic stated that its annualized revenue had surpassed $30 billion, more than tripling from approximately $9 billion at the end of 2025. Subsequent Q1 global LLM market data released by Counterpoint Research showed Anthropic topping the global rankings with a 31.4% revenue share, surpassing OpenAI's 29.0%.
More impactful is revenue efficiency: Anthropic's monthly active users are only 134 million, about one-seventh of OpenAI's, but its average monthly revenue per user is about $16, more than 7 times that of OpenAI and over 160 times that of Meta.
Generating more revenue with fewer users indicates that Anthropic is capturing the layer of productivity users with the highest willingness to pay and the greatest usage intensity.
By the time of the May developer conference, CEO Dario Amodei provided updated figures: annualized revenue had surged to over $44 billion, inference gross margin had climbed from 38% a year ago to over 70%, enterprise clients spending over $1 million annually had grown from just over a dozen two years ago to more than 1,000, including 8 of the Fortune 10.
The speed of revenue growth exceeded Amodei's own expectations. He said at the conference, "I hope 80x growth doesn't continue; that's crazy and too hard to manage."
The core product driving this growth curve is Claude Code. This programming agent tool had an annual recurring revenue exceeding $2.5 billion in early 2026, with a market share of about 54% among similar tools. Once enterprises deeply integrate Claude Code into their development processes, migration costs become extremely high, creating a certain stickiness barrier.
But it was precisely at this moment that OpenAI struck.
In mid-April, a four-page internal memo from OpenAI's Chief Revenue Officer Denise Dresser was published in full, directly accusing Anthropic of using the "gross method" for revenue recognition. Specifically, when enterprise clients purchase Claude services through AWS, Google Cloud, or Azure, Anthropic records the full amount paid by the client as its own revenue, including the portion that needs to be shared with the cloud service providers as channel fees. OpenAI, on the other hand, uses the "net method," booking only the net revenue after deducting Microsoft's share.
According to Dresser's calculation, if standardized to a net basis, Anthropic's claimed $30 billion annualized revenue should be approximately $22 billion, lower than OpenAI's $25 billion during the same period. "Their story is built on fear and constraints," Dresser wrote in the memo.
Media analysis pointed out that both accounting methods are legal under US GAAP, with the key lying in the company's role positioning in the transaction. Anthropic's stance is that Claude is the core product, and cloud platforms are merely distribution channels, justifying the gross method. However, the issue is that Anthropic's relationship with cloud platforms is not just distribution—Amazon and Google are simultaneously its shareholders, compute suppliers, and distribution partners. When funds circulate within this triangular relationship, distinguishing between "channel fees" and "investment returns" becomes a conundrum in itself.
Bank of America provided an estimate: it is expected that in 2026, the total channel commissions paid by Anthropic to AWS and Google will reach $6.4 billion, more than doubling from $1.9 billion in 2025. Under the gross method, this $6.4 billion is counted as part of revenue; under the net method, it would never appear in revenue at all.
OpenAI's choice to strike at this juncture is not just an accounting debate. Both companies are advancing towards an IPO. Once the S-1 registration statement is submitted to the SEC, regulatory bodies will force both sides to recalculate revenue under a unified framework.
Khosla Ventures partner Ethan Choi, in a previous interview with Forbes, got to the heart of the matter: "If they both IPO in the next few quarters, I'm not sure the SEC will allow two companies to use different accounting treatments for essentially the same type of revenue."
This dispute over $8 billion in accounts is both an offensive and defensive battle over valuation and a stress test that both companies must face as they rush towards the public market.
The greatest weakness of the Suzerain State is not its growth speed, but that its growth numbers have not yet been audited. Every line in the S-1 registration statement related to revenue quality, depreciation periods, and related-party transactions could become a measuring stick in the hands of skeptics.
03
Discourse Power Is Handed Over
The Suzerain State becomes a Suzerain State not because of what it says itself, but because the empires that support it are reorganizing themselves according to its path.
Amazon is betting the next decade of its Trainium chips on Claude; Google is handing over its most advanced TPU production capacity; Microsoft is deeply integrating Claude into Azure while renegotiating its cooperation agreement with OpenAI; SpaceX has written preliminary intentions to cooperate with Anthropic in developing multi-gigawatt orbital AI compute power.
When supporters successively hand over the configuration rights of their core strategic assets to the same entity, that entity no longer needs to compete for discourse power, because discourse power is handed to it.
The FTC warned about this structure as early as January 2025 in its AI cooperation report: when cloud credits, compute commitments, equity, and revenue-sharing terms are nested together, they can "shape competition, switching incentives, and access to commercially sensitive information." More than a year later, this warning almost accurately predicted Anthropic's support system.
And Anthropic is converting this structural advantage into ecosystem control. Claude is the only frontier model simultaneously running on the three major mainstream cloud platforms: AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure Foundry.
This cross-cloud capability means enterprise clients are not locked into a specific model by choosing a particular cloud provider. Conversely, if a cloud provider rejects Anthropic, it risks being abandoned by enterprise clients. The cross-cloud portability of models has inverted the supply-demand relationship. In enterprise AI spending, Anthropic's relative share has jumped from about 10% in early 2025 to over 65% in February 2026.
Another variable with structural significance is Claude Code. This programming tool already accounts for approximately 4% of all public commits on GitHub globally. Code is not an ordinary consumer category; it is the underlying grammar of the modern digital economy. Whoever controls the entry point for code generation holds the power to define the mode of software production. When developers become accustomed to Claude Code's workflow, it's not just Anthropic's revenue that grows; the center of gravity of the entire software development paradigm also shifts towards it.
OpenAI is not without counterattacks. The Dresser memo claimed that OpenAI expects to have 30 gigawatts of computing power by 2030, while Anthropic will have only 7 to 8 gigawatts by the end of 2027. However, the significance of computing power lies not in total volume but in timing. Colossus 1 is currently the only option that can be delivered immediately, while OpenAI's long-term compute commitments will take years to materialize. Enterprise clients' migration decisions happen now, not in 2030.
Meanwhile, OpenAI is advancing its own multi-cloud strategy and independent compute layout, with its exclusivity relationship with Microsoft already loosening. Anthropic's multi-cloud architecture, on the other hand, was designed from the outset; it doesn't need to break free from any constraints because it never bound itself tightly to any single giant.
The cooperation agreement between Microsoft and OpenAI allows OpenAI to serve clients through any cloud provider, while Microsoft's license to OpenAI's models and product IP will extend until 2032. The essence of this clause is that both companies are preparing for the eventual separation, and Anthropic is reaping the strategic space created by this rift.
04
The Suzerain State Is a Timed Game
The Suzerain State is a temporary configuration.
Anthropic depends on its supporters for computing power, and the supporters depend on Anthropic for models. They need each other and are wary of each other.
Musk retained a clause in the SpaceX agreement: if Anthropic's AI acts in a way harmful to humanity, SpaceX has the right to reclaim the computing power. This clause looks like a safety statement but is actually a substantive control arrangement, a microcosm of the bargaining within the entire Suzerain State system.
But at least in the current time window, the party with the most supporters is pushing the industry forward according to its will. In five years, Anthropic has transformed from a research lab into the central dispatcher of AI infrastructure.
If the $950 billion valuation financing is completed by the end of May and a subsequent IPO truly happens, this story will enter a completely different chapter. Previous valuations were a pricing game in the private market; the IPO prospectus will be the first real external scrutiny.
At that moment, the $8 billion accounting dispute in the Dresser memo will no longer be one company attacking another but a formal inquiry on the SEC's desk. The gains for Amazon and Google will also transform from paper appreciation into related-party transactions that must be explained item by item in audit reports.
For the Suzerain State to be sustained long-term, it needs to convert its nourishment into self-sustaining capability. Anthropic's inference gross margin increasing from 38% to over 70% is the most solid foundation in its business model narrative. However, if the difference between gross and net methods casts doubt on revenue quality, or if computing costs remain high and growth slows without rapidly reducing losses, this narrative faces repricing.
Yet, even with all these uncertainties laid out on the table, one fact is irreversible: the power within the AI industry is being redistributed along the lines of the compute supply chain, cloud distribution channels, and developer ecosystems. And at the current juncture, the coordinate closest to the center of this power is called Anthropic.
This article is from WeChat public account "Silicon Starlight", author: Yuan Tai
