From AI Lab to Laser: What Anthropic's S-1 Reveals About the Optical Bottleneck

With $ANT filing, the AI stack is public from top to bottom for the first time. Six links from the lab's tokens to the lasers that connect its compute.

In this piece:

1. The missing layer: why ANT going public “completes the AI stack”
2. Why AI is still early: adoption, inference and the super cycle
3. Six links from AI lab to laser: tokens, dollars, gigawatts, the copper wall, light and lasers
4. Bubble or build-out? What has to hold, and what would break it
5. The stack, layer by layer: leaders and challengers from AI labs to materials
6. The bear case: the risks in detail
7. From AI lab to laser: the chain in one paragraph

On my previous AI infrastructure series I talked about the AI Stack. Part 1 and Part 2 cover the compute and interconnect layers in more depth.

The missing layer

When I mapped the AI compute stack in Part 1, it started with the hyperscalers. That was where the public data began. But the hyperscalers are not where the demand begins. Above them sits the actual fundamental AI Layer: the AI labs that turn compute into intelligence and sell it by the token.

Until this week, that top layer was private. Investors could see Google's capex, Nvidia's revenue and Lumentum's laser shipments, but not the buyer whose demand was pulling all of it. On October 26, Anthropic filed its S-1 to list on Nasdaq as ANT. For the first time, the whole stack is public, from the lab to the laser:

This matters all the way down to the photonics and lasers, because the S-1 lets us follow demand down the stack with numbers, not assumptions. And read as a photonics document, the filing is revealing. In its glossary, "compute" means not only accelerators but also memory, networking, storage and power. By Anthropic's own definition, every compute dollar includes the links between the chips. Its risk factors add that demand for compute has historically exceeded available capacity.

The S-1 at a glance, read as a map of the stack:

• Demand: $16.3 billion of revenue in the first half of 2026, 80% of it billed per token.
• Compute: $8.9 billion spent on compute and infrastructure in the same half year, 55% of revenue.
• Commitments: roughly $780–800 billion of compute agreements running out to 2045, most of them non-cancelable.
• Architecture: one shared compute pool across four chip families: Trainium, TPUs, Nvidia and AMD.
• Constraint: demand for compute has exceeded supply, and Anthropic expects that to last.

This piece is not a review of the IPO. It uses the S-1 as a lens, because it is the first time we can follow AI demand from a lab's revenue all the way down to the lasers.

When LAZR added Anthropic in July, ETF.com called it a puzzling fit, because Anthropic is a customer of photonics, not a supplier. That is true, and it is a fair point. However, Anthropic is on top of the stack and this piece walks down the stack from the top, in six links: tokens, dollars, gigawatts, racks, light and lasers, hence the connection is very real.

Why AI is still early

It is easy to look at Anthropic's growth and assume the curve must be near its top. The S-1 suggests the opposite. Measured by how deeply AI is actually used, we are still near the bottom.

Adoption is wide but shallow. More than half of the Fortune 500 already spend over $100,000 a year with Anthropic. Yet on average that spending is less than 0.1% of those companies' total costs. Most organizations have started using AI. Very few have rebuilt how they work around it.

Most businesses have not started at all. The US Census Bureau's business survey found that only about 20% of US companies used AI in any part of their business as of May 2026 (Census Bureau). That is up from under 4% in September 2023, so adoption has grown fivefold, yet four in five businesses still do not use it. Workers are ahead of their employers: about 45% of employed US adults already use generative AI at work. Where people lead, companies tend to follow.

The work is changing shape. The first wave of AI answered questions. The current wave does tasks. Anthropic's stated goal is AI that works on its own for hours, days or weeks, and the S-1 says Claude's success rate on hard engineering problems jumped by about 50 percentage points in nine months. Each step from answering, to doing, to working for days multiplies the compute behind every user.

AI is moving from the lab into production. Training builds a model once. Inference is the model doing work, every time someone uses it, and it never stops. Deloitte estimates inference was about a third of all AI compute in 2023, half in 2025, and will be roughly two-thirds in 2026 (Deloitte). Gartner expects inference spending in AI cloud infrastructure to overtake training this year. That shift is the clearest sign AI has become a working tool rather than an experiment. It also means compute demand now grows with usage, not just with each new model, and reasoning and agentic models make every request heavier. Deloitte's own conclusion is that the next phase will demand more computing power, not less.

The market is enormous and barely touched. The S-1 puts the value of global knowledge work at more than $40 trillion a year, and counts more than a billion knowledge workers by 2028. Coding was the first job AI changed at scale. Office work, research, finance, law and science are next, and Anthropic says its office product is spreading faster than its coding product did.

Supply is the limit, not demand. Anthropic says demand for compute has exceeded what it can get, and expects that to last for the foreseeable future. That is unusual for a technology at this scale, and it is the clearest sign the cycle is still early.

Even the experts are only now reorganizing. On October 26, Microsoft said it would offer voluntary buyouts to about 60,000 employees, roughly a quarter of its workforce, to retool for the AI era (Bloomberg). Microsoft builds AI infrastructure, sells AI products and is one of Anthropic's compute partners. If a company that close to the technology is only now reshaping how it works, most of the economy has barely begun. Anthropic shows where that leads: revenue per employee of more than $6.6 million, up from $1.4 million a year earlier. As that kind of efficiency spreads, each organization runs more of its work on AI, and every one of those tasks runs on compute.

The build-out has only just started. In Part 1 we looked at Goldman Sachs' estimate of $7.6 trillion in AI capex through 2031. Global data center capacity is expected to roughly double by 2030. Anthropic alone has signed compute agreements running to 2045.

The long-term demand case: 10,000x. At Ayar Labs' CPO Summit this autumn, Pat Gelsinger, former Intel CEO and now an Ayar Labs board member, argued that the world will need roughly 10,000 times more AI than it uses today (EE Times).

That is a compound of better models, wider deployment and deeper enterprise use, not a forecast of 10,000 times more data centers. His own point is that power cannot scale anywhere near that far, so AI economics must improve by four to five orders of magnitude, and the network becomes the binding constraint. In his words, "the network is the AI." Even $7.6 trillion of capex does not get us to 10,000x. Efficiency has to close the gap, and much of that efficiency has to come from moving data with light instead of copper.

Put together, this is what an AI super cycle looks like from the inside: usage is growing exponentially, every user consumes more compute each year, and the physical stack underneath is being built years ahead of demand. The rest of this piece follows that demand down the stack, to the layer where it turns into light.

Link 1: Revenue is tokens

Anthropic sells intelligence by the token. Consumption pricing, metered in tokens and priced by the model the customer picks, was 80% of first-half 2026 revenue.

That meter is running fast. Revenue went from $41 million in 2023 to $386 million in 2024 and $4.6 billion in 2025. The first half of 2026 alone brought $16.3 billion, fourteen times the same period a year earlier.

The S-1 also defines the run-rate figures that have filled the headlines: the last 28 days of consumption fees multiplied by 13, plus monthly subscriptions multiplied by 12. A run rate is a snapshot, not a forecast. But the customer base behind it is broad: more than 11,900 customers spending $100,000 or more a year, about 250 spending $10 million or more, and net revenue retention of roughly 900%.

Press reconstructions put Q3 revenue at about $16.9 billion, roughly 47% above Q2. Growth is slowing in percentage terms, after a 144% jump the quarter before. It is still extraordinary.

Link 2: Tokens grow faster than dollars

If revenue is tokens, the more important curve is the token count itself. Two charts published this month show it.

Bank of America, "Token Costs and Token Usage"

Bank of America's tracker shows token volumes across US and Chinese labs rising by orders of magnitude since early 2025, with Anthropic now the largest share. Over the same period, the average price per million tokens roughly halved, with most of the drop since August

Ramp's index of business API spend shows the same thing from the customer side: token usage up roughly 170-fold since January 2025, and still accelerating into October.

The S-1 explains the mechanism in Anthropic's own words. Claude Opus 5.5 cut customer cost per task by about 40% compared with Claude Opus 5, and the filing says lower cost per task drives greater usage. New models also keep resetting demand: models released in the six months to September drove more than 80% of Q3 revenue. Agentic work multiplies usage too: in Q3, Claude Enterprise users without Claude Code spent under $75 a month on average, while the top 5% of Claude Code users spent about 20 times that.

The filing also shows that dollars of compute and units of compute are not the same thing. In Q2, revenue grew 1,368% year on year while compute and infrastructure expense grew 300%. Part of that gap is what Anthropic calls compute multipliers: algorithmic improvements that get more output from the same hardware. Physical compute grows faster than the compute bill, and tokens grow faster still.

As Pat Gelsinger put it at Ayar Labs' CPO Summit, the world will need roughly 10,000 times more AI than it uses today. Token volume can grow a hundredfold while revenue grows tenfold and the price per token falls. That only works if the cost of producing a token falls just as fast, every generation. Hold on to that. It is the reason this chain ends in light.

Link 3: Dollars become gigawatts