My Story with Anthropic

Some stories you only notice you were living once they've already become history. Mine with Anthropic is one of those. I didn't set out to have a "story with Anthropic" — I set out to build an inference engine. But across three and a half years I had the rare seat of both collaborator and observer: close enough to debug a launch at 3 a.m. with their engineers, and far enough back to watch a company go from "Claude? what's that?" to a household name worth nearly a trillion dollars.

This post is that story, told along its timeline. I've drawn the arc as a diagram first, then walked through it chapter by chapter — partly a record of the collaboration, partly an honest log of what I observed from the inside as the frontier moved.

The Arc, at a Glance

The whole story compresses into a single rail — from the day I became engineer No. 2 on a confidential project, to a future milestone I've left blank on purpose:

My Anthropic timeline — the hollow, dashed node is the future, not yet lived.

---

The Inference team was newly funded when I joined in late 2023. I was the first engineer hired, alongside one L6 engineer who'd come over from Lex — which made me, in effect, engineer No. 2. The mandate was privileged and confidential: build the inference engine for Anthropic's models on Amazon Bedrock. For nearly a year our work was shrouded in secrecy — for partner-protection and IP reasons (on Anthropic's behalf), even other Bedrock teammates couldn't access our codebase, documents, or channels. For a long stretch, owing to org restructuring, I was even the only L6 senior engineer on the team.

From the very start, ours was the only team working directly with Anthropic. I still remember an L7 colleague Slack-messaging me early on: "Andy, good to meet you. I've heard your name mentioned even by [the co-founder of Anthropic], when we were working on this project." It's a humbling thing, to learn your name is being said in rooms you've never been in.

This was, once again, a 0→1 startup — clean code, high visibility, fast pace — except this one sat inside what would become the most important business at the company.

We partnered directly with Anthropic — including co-founder Ben Mann — with two-plus working sessions a week and a monthly happy hour to celebrate each new model release. That work touched ideas that have since become standards across the open-source inference community — learn more in my other blog post →

Through 2024 I led almost all of Claude's Day-0 public releases on Bedrock. The cadence was relentless — roughly one or two major launches a month — full story in Bedrock 1 Year →

• April — Claude 3 Opus on Amazon Bedrock.
• May — Claude 3 Sonnet and Haiku in Frankfurt.
• July — Claude 3.5 Sonnet.
• October — upgraded Claude 3.5 Sonnet v2 with Computer Use, Anthropic's first agentic-AI feature.
• November — Claude 3.5 Haiku GA; Anthropic + Palantir bringing Claude to the U.S. government cloud.
• December — prompt caching, latency-optimized inference, and Claude 3.5 Sonnet in the AWS Top Secret cloud (re:Invent).

I also delivered the first Claude model on AWS Trainium/Inferentia, partnering with James Bradbury (Anthropic's Head of Compute) and the Annapurna teams. There's a serendipity I still smile at: through my Linux Foundation community work I'd been invited to San Francisco on the PyTorchCon Program Committee, where I sat on a panel alongside James. The following week, back in Seattle, I found myself working with him full-time on Claude on Trainium. You meet someone outside of work, and a week later you're shipping a frontier launch together.

I vividly recall several late nights where both teams worked until sunrise to debug a launch — and the happy hours afterward in Seattle's South Lake Union that made the exhaustion worth it. This was also the first time I felt how intensely Amazon's most senior leadership cares about a business: we took tickets directly from customer CEOs, and every Sev-2 review pulled in L10 leaders — our VP and a Distinguished Engineer — alongside L8 Senior Principal Engineers.

And here is the part that dates this chapter most precisely: when I told people at CVPR 2024 that I worked on the Claude model, I'd often get a blank look. They didn't know what it was. Anthropic was, at that point, a frontier lab valued at roughly $18 billion — serious money, but not yet a name your family would recognize.

2025 was the year the wider industry caught up to what we'd been living. Anthropic became widely recognized across the field, and in May 2025 the company released Claude Code — the tool that, more than any single model, made developers everywhere fluent in the name "Claude." SemiAnalysis would later estimate Bedrock as a multi-billion-dollar run-rate business with the vast majority of its customers (80–90%+) on Anthropic models.

"SemiAnalysis believes Bedrock is a $5.5B run rate business today with the vast majority of customers (80-90%+) using Anthropic models."

After shipping my last 3P task in Q1 2025, I chose to step sideways. Bedrock Inference was organized into 3P (Claude and other closed models), 2P (open-weights — Gemma, gpt-oss, Qwen, DeepSeek), and 1P (Amazon's own Nova and Titan). I moved from 3P into 2P, beginning my research line in model optimization (speculative decoding) and customization — the work I still lead today. I left the Claude program, but I never stopped watching it.

By 2026 the blank looks were gone. Regular users worldwide — including people in China — had become familiar with Claude and were learning, on their own, how to use it well. The thing a handful of us had built behind a badge-locked door had quietly become part of everyday life.

There is one number from this year I find hard to look at without feeling something. In 2Q26, SemiAnalysis put more than 40% of Anthropic's ARR through indirect channels: Amazon Bedrock, Microsoft Foundry, and Gemini Agent Enterprise, with Bedrock the largest of the three. Their own chart makes the second half of that plain: Bedrock grew from roughly $150M in a quarter in 1Q25 to a forecast $9.3B in 4Q26, and about 40% of Anthropic's API revenue now arrives through it. The serverless endpoint a handful of us stood up behind a locked door, the one I had to explain at CVPR in 2024, is now one of the main doors that revenue walks through.

2Q26  ·  where the revenue arrives

Four in every ten of Anthropic's ARR dollars now arrive through someone else's console.

Anthropic ARR by channel, 2Q26

Indirect 40%+

Direct ~60%

• Indirect: Amazon Bedrock, Microsoft Foundry, Gemini Agent Enterprise. The customer pays a cloud bill, and Anthropic is the seller of record.
• Direct: Anthropic's own API, the Claude apps and subscriptions, and enterprise contracts signed directly.

Bedrock revenue, and how much of Anthropic's API runs through it

swipe to pan the chart →

• Amazon Bedrock revenue, $M per quarter
• Share of Anthropic's API revenue via Bedrock

The 1Q26 bar, roughly $1.37B in a quarter, is the $5.5B run rate quoted above. Note that this 40% and the 40% in the panel above are different measurements that happen to land on the same number: this one is Bedrock's share of Anthropic's API revenue, the one above is all indirect channels' share of total ARR.

How one Bedrock dollar splits

CustomerPays a single AWS bill, inside an account whose security review and contract are already done.

AnthropicIs the seller of record, so it books the full token spend as gross ARR.

AWSKeeps an IaaS fee plus a revenue share. Anthropic books that in sales and marketing: it trims lab EBIT, but costs far less than the enterprise sales force it replaces.

I want to be careful about what that chart does and does not say. It is not my number. It is Anthropic's revenue, and the channel carries vastly more than the code any one engineer wrote into it. But the shape of it is the thing: a frontier lab now reaches a large share of its customers through a console someone else operates, and being the seller of record on that console was a structure Anthropic did first, with AWS. In 2024 we were arguing about token accounting and cold-start latency on a service most people had never heard of. It turns out we were building a distribution channel.

I watched Anthropic itself grow up close — from a team of fewer than 200 people to what it is today, its valuation climbing from roughly $18B in 2024 to ~$965B by mid-2026. More than a 50× jump in two years, making it the world's most valuable startup. Some of the most valuable parts of this chapter weren't in any codebase at all — they were the people the work put me next to: engineers, founders, researchers, and investors I now count as part of my network.

If there's one thing that stayed constant across every person I worked with at Anthropic, model release after model release, it's this: AI safety is a root philosophy, not a checkbox. As early as 2023, Anthropic published its Responsible Scaling Policy, defining a framework of AI Safety Levels (ASL) — safety, security, and operational standards that scale up as a model's capability, and its potential for catastrophic risk, scales up.

CEO Dario Amodei has said as much himself, in different settings and in his own words, about how much weight trust and safety carries in the company's decisions. I saw that weight directly — I was one of the engineers who helped build, alongside Anthropic, multiple safety-system ecosystems for trust-and-safety control of increasingly powerful models, for every model release we shipped together. There was very little room for compromise in that process. Of every model provider I've worked with or watched, I have never seen one take it as seriously as Anthropic does.

That commitment extends into research, not just policy. Anthropic's recent interpretability work — Global Workspace and Tracing the Thoughts of a Language Model — is, to me, part of the same lineage: understanding why a model does what it does is a prerequisite for trusting it at scale. That kind of work is a safeguard for the whole trajectory toward AGI. Anthropic has done the work to lay that foundation solidly, and it's what lets the rest of us keep building on top of it with confidence.

Every arc this steep eventually reaches a public-market milestone. As I write this, Anthropic's IPO is still ahead of us — a date I genuinely don't know yet, and so one I'm deliberately leaving blank. When it lands, this is where I'll mark it.

When I think back over this whole timeline, what stays with me isn't any single launch. It's the strange privilege of the double seat — building the thing and watching it become history at the same time. In 2024 I had to explain what Claude was; in 2026 I no longer have to explain anything. The rest of that story, Anthropic gets to write. I just got to be there for the beginning.