Anthropic said in April that run-rate revenue exceeded $30 billion, up from about $9 billion at the end of 2025. Its May Series H announcement said the figure had crossed $47 billion earlier that month. It also said customers spending more than $1 million annualized had risen from more than 500 in February to more than 1,000.
Over the same period, Anthropic disclosed up to 5GW of new Amazon capacity, 5GW of next-generation TPU capacity with Google and Broadcom, and access to GPU capacity at SpaceX Colossus 1 and 2. Amazon remains its primary cloud and training partner. Adding the numbers produces a portfolio above ten gigawatts, not capacity simultaneously available today.
Run-rate is neither quarterly revenue nor GAAP. It usually annualizes a recent activity level and can magnify exit velocity in a rapid-growth period. Gigawatts may be future ceilings, phased delivery and different hardware platforms. Both sets of figures require conversion before they are compared.
$47 billion must be decomposed into customer, price and retention
The doubling of large enterprise accounts indicates breadth beyond one customer, but outsiders still lack top-ten concentration, API-versus-subscription mix, discounts, compute-resale share and net retention. Run-rate supports demand directionally; it does not replace quarterly P&L and cash flow.
Claude Code and long-running agents may increase workload depth, while better success rates may reduce retries. Revenue growth cannot be mapped directly to tokens and then to GPUs or MW. Pricing, caching, model routing, inference efficiency and workload mix change each conversion.
| Metric | Disclosure | Basis and boundary |
|---|---|---|
| End-2025 run-rate | ~$9B | Company-disclosed annualized measure. |
| April 2026 | >$30B | Disclosed with the Google-Broadcom partnership. |
| May 2026 | >$47B | Series H disclosure; not quarterly revenue or profit. |
| Large enterprise accounts | 1,000+ | >$1M-plus annualized spend; concentration remains undisclosed. |
Multi-platform compute is supply and workload insurance, not duplication
AWS Trainium, Google TPU and NVIDIA GPU differ in more than performance. A portfolio reduces vendor and regional concentration and distributes training, inference and customer-cloud delivery. It also requires compiler, kernel, network, scheduler and operational work across platforms.
The Amazon and Google-Broadcom 5GW figures are future arrangements. SpaceX offers access to existing Colossus clusters. They can overlap or complement one another across years, regions and workloads. More than ten gigawatts is a portfolio ceiling and roadmap, not today's simultaneously loaded facilities.
Primary training partner
Up to 5GW links Trainium to a long cloud relationship, with delivery phased by project.
Next-generation TPU
The 5GW plan begins from 2027 and emphasizes price-performance and platform diversity.
GPU access
Colossus offers a distinct and potentially faster path, while price and contracted capacity are not fully disclosed.
Revenue covers compute commitments through a four-layer schedule
Layer one is customer revenue: contract, use and collection. Layer two is cloud and silicon commitments: minimum purchases, prepayments, leases and cancellation. Layer three is capacity delivery: land, power, building, equipment and acceptance. Layer four is finance: equity, strategic investment, debt and partner-owned assets.
The $65 billion Series H expands liquidity and includes hyperscaler strategic capital. It does not mean every future compute payment is funded. If capacity commitments lead collections, Anthropic still needs capital markets or suppliers to carry timing risk.
A useful but currently unobservable metric is compute coverage ratio: sustainable gross cash generation divided by non-cancellable compute payments over the next twelve months. Without contracts and costs, only ranges are defensible.
- 01Customer run-rate
Annualized recent activity must be tested for concentration, retention and collection.
- 02Contract commitment
Separate maximum capacity, minimum payment, option and cancellation.
- 03Available capacity
Only energized and accepted clusters can serve revenue.
- 04Cash coverage
Gross cash and financing together fund prepayments, leases and service fees.
Both the financing-bottleneck and circular-finance theses need falsification
The bull case says high-value enterprise agents generate revenue far above annual compute and operating cost per GW. The rational strategy is maximum finance and construction speed; capital, not end demand, limits growth, and hardware diversity reduces supply risk.
The bear case says run-rate overstates durable revenue, customer concentration or pricing pressure compresses margin, and giant compute commitments depend on cloud-provider investment and supplier finance. Higher utilization risk or funding cost then widens the cash gap quickly.
The tests are concrete: quarterly revenue and collections, concentration, live MW by platform, non-cancellable payments, inference margin and renewal. As those appear, narrative can become a model.
IDC ATLAS VIEWAnthropic's strongest signal is not a single $47 billion or 5GW figure but the simultaneous acceleration of revenue, funding and multi-platform capacity. The most important gap is equally clear: future compute must be covered by durable gross cash, not only the next financing round.