Anthropic projects $70b revenue, $17b cash flow by 2028

Anthropic expects to generate up to US$70 billion in revenue and US$17 billion in cash flow by 2028.

The AI startup’s revenue from AI model API access is projected to reach US$3.8 billion this year, driven by strong enterprise demand.

It recently partnered with Microsoft and expanded its collaboration with Salesforce, and plans to roll out its Claude AI assistant to employees at Deloitte and Cognizant.

The company launched new models, Claude Sonnet 4.5 and Claude Haiku 4.5, targeting businesses deploying AI at scale.

Anthropic last raised US$13 billion in September at a US$170 billion valuation and could seek up to US$400 billion in future fundraising.

Its gross profit margins are expected to improve from negative 94% last year to 50% this year and 77% in 2028.

Rival OpenAI projects US$13 billion in revenue this year but continues to face losses due to infrastructure investments.

🔗 Source: TechCrunch

🧠 Food for thought

Implications, context, and why it matters.

Anthropic’s margin story depends on cloud contract terms we don’t have

2024 AWS spend hit $1.35 billion on an estimated $400-600 million in revenue 1. Through September 2025, AWS costs reached $2.66 billion on $2.55 billion in revenue 1.
Anthropic splits training across Trainium (Amazon’s custom AI training chip), Google Cloud TPUs (Tensor Processing Units; Google’s AI accelerators; up to 1 million units expected in 2026), and NVIDIA GPUs (Graphics Processing Units) 2 3. Terms on minimums and revenue share remain undisclosed. So do exclusivity and access guarantees. That leaves 77% margins by 2028 unproven.
Priority Service Tiers arrived in May 2025, with higher prices for capacity or lower latency. After the change, Cursor’s bill rose from $6.2 million to $12.6 million in one month on Amazon’s cloud 1.

Multi-cloud Claude rollout opens room for third‑party cost tools

Claude runs on AWS Bedrock (Amazon’s managed service for hosting third-party AI models) and Google Vertex AI (Google’s managed AI platform). Bedrock lists $0.008 per 1,000 input tokens and $0.024 per 1,000 output tokens for some models 4. Prices differ by region and platform 5 6 7.
FinOps platforms and broker proxies can cut spend by routing across clouds. No shared layer handles price, latency, or availability in real time.
Building this needs up‑to‑date cross‑cloud Claude prices, regional maps, and latency tests. Bedrock uses token pricing 4 and Vertex uses character pricing 6. The data is scattered, yet it could unlock savings for 300,000+ businesses using Claude 2 7.