The AI value chain, measured from the silicon up

Start where the numbers are audited and the first thing you find is that the forecasts are already behind. Omdia projected a US$286bn market for AI chips in 2030. Reported sales passed that figure in 2026, four years early. That is the argument for deriving this chain from filings instead of from anyone's growth rate.

Nobody reports chips, they report dollars, so units are solved from revenue. NVIDIA's datacenter segment billed US$162.4bn in its fiscal 2025. At about US$42,000 per chip, that is 3.87M chips sold in the year. For 2026 the four buckets (NVIDIA, AMD, the custom chips cloud firms design for themselves, and China) sum to US$381bn of accelerators. Then comes networking: the switches, cables and cards that wire thousands of chips into one machine. NVIDIA reports it at 24.5% of accelerator revenue, which lifts the total to US$474bn, within a few percent of what the four report to the SEC.

The industry frames 2026 as a growth story. It is a price story. AI-server shipments grow 28% while silicon revenue climbs about 90%, because the average rack price tripled in two generations (GB200 near US$3.0M, GB300 near US$4.0M, Vera Rubin up to US$8.8M). NVIDIA is already touting US$1 trillion of Blackwell and Rubin orders through 2027, more than Switzerland's GDP.

So 2026 is as much a price cycle as a unit cycle. The number to watch each quarter is datacenter revenue against the price-per-chip assumption, not the unit-count headlines that double-count multi-die packages.

A 2023 GPU and a 2026 rack are not the same unit, so every chip generation is converted into H100-equivalents using public benchmarks. One H100-equivalent means as much computing as a single 2023 H100 chip, which lets silicon from different years add up in one honest unit.

The conversion indices are measured, not assumed. The H100 sits at 1.00. The B200 at 2.90 (SemiAnalysis clocks its cost from US$0.49 to US$0.17 per million tokens). The GB300 at 4.50 (MLPerf shows 45% more than the GB200). The unit tracks work actually delivered, not the peak speed on the spec sheet.

This is also where the loudest counting error hides. Headlines count 6M GPUs where the filings describe 3M packages, because the newer parts carry two chips each. Halving the count halves the implied price per chip, and that error then runs through every estimate below it. The model counts packages, exactly as the filings do.

Chips sold are not chips running, so the fleet is built as a sum with memory. Each year's purchases come in; silicon past its service life goes out. What comes in is the shipment count from link 01, converted at link 02's indices. The 13.08M packages sold in 2026 are worth 35.0M H100-equivalents, because the average part of that year does 2.7 times the work of a 2023 H100. The second chart below shows those annual additions. Add the cohorts (2.1M from 2023, 5.4M from 2024, 16.2M from 2025 and 35.0M from 2026) and 58.7M H100-equivalents are standing by the end of 2026.

Not all of that runs a full year. A chip delivered in March cannot work twelve months, so new silicon is assumed to run half the year in its first year. The working fleet in 2026 is therefore 23.7M carried from end-2025 plus half of the 35.0M just delivered: 41.2M H100-equivalents actually running. How long a chip lasts is lever A7.

That lever is not a detail. Service life is the number-one variable of total cost, and the hyperscalers (Microsoft, Alphabet, Amazon and Meta, who rent out most of the world's computing) have already stretched it from 3 to 5-6 years in their accounts. Every extra year cuts the annual cost without moving a dollar of spending. It belongs on the dashboard as a lever, not buried in a footnote.

A token is AI's unit of work, roughly a word read or written. Rather than trust what the chips could do on paper, the fleet's output is calibrated backwards against the one volume anyone has disclosed: Google's 3.2 quadrillion tokens per month. The model solves how much each chip produces instead of assuming it.

The arithmetic is direct. Take the global rate for 2026, 6.0 quadrillion tokens a month, multiply by twelve, and divide by the 41.2M working fleet. That gives 1.75 billion tokens per chip per year, solved rather than assumed. But the 6.0 is my own estimate, not a disclosure. It is built bottom up from what is public, mostly Google's 3.2 quadrillion a month and OpenAI's reported throughput, and that bottom-up work gives a range of 5.5 to 7.7 quadrillion. Choosing 6.0 is the same as saying Google runs 53% of the world's tokens. That one share is what the whole step rests on.

The sanity check is what makes the calibration believable. 1.75 billion tokens a year is 55 tokens per second, sustained. That is about 3% of what an H100 can do at full tilt. Low, and correctly so, because part of the fleet is training models rather than answering users. Measured efficiency gains run at 41% a year (SemiAnalysis) against the 60-70% sometimes claimed, while MLPerf shows 45% per hardware generation. I believe the measurement, not the press release.

Two conclusions follow, and they matter more than the calibration itself. Chips do keep getting cheaper per token, at a measured 41% a year, but the fleet grows faster than that. So almost every extra token comes from buying more silicon, not from better silicon. And at 3% of its paper peak, the fleet is nowhere near the ceiling of what it could print. Token supply is not what limits this industry. What each token sells for is.

The one genuine cross-check is a rival estimate, and it does not settle the question. Goldman puts the global rate at 5.0 quadrillion a month, about 20% below mine. On their number Google runs 64% of the world's tokens; on mine it runs 53%. Neither share is absurd, so I carry the gap as uncertainty rather than argue it away. A second lab disclosing its volume would settle it. What is not in doubt is the direction: OpenAI's API went from 6 to over 15 billion tokens per minute in five months. Agents change the scale outright, since one coding-agent session burns 100 to 3,500 times the tokens of a chat, already about US$13 per developer per active day.

Not every token that runs is a token that sells. The 55% that does (lever B3) is not a guess, it is a blend of two measured extremes. Google bills only 26% of its tokens through its API, the paid interface other companies plug into. The rest of its inference feeds its own products, free tiers and test traffic. OpenAI and Anthropic sit near the other end, close to 100%, since almost everything they run is sold. Weighted by size, the two ends give about 55%.

Price is anchored to booked revenue, not to a run-rate (an annualised figure built by multiplying one good month by twelve). US$55bn of booked revenue in 2026 over 39.6 quadrillion billable tokens gives US$1.40 per million tokens. That price (lever B4) is the single most uncertain input in the chain, for a plain reason: nobody publishes it. No lab reports revenue divided by tokens, so the figure has to be assembled from two estimates at once, and both of them move. It is also not a list price. It blends expensive reasoning models with cheap ones, volume discounts and cached traffic, and that mix shifts every quarter.

Two forces pull it down, and the market routinely misreads both. Run-rate flatters the number. Anthropic closed 2025 at a 'US$9bn run-rate' but booked US$4.5bn of actual revenue for the year, a 2x overstatement that inflates the implied price with it. And the cheap end of the market is no longer marginal. OpenRouter is a service that routes developer traffic to whichever model the developer picks, and it publishes its mix. Open-weight and Chinese models went from under 2% of that traffic to over 45% in twelve months. Every user who moves there keeps working and pays less, which drags the blended price down without any lab cutting a published price.

It does not only fall, which is what makes it so hard to pin down. Anthropic raised Sonnet 5 by 50%, and its new tokenizer emits about 30% more tokens for the same text. That is a price increase that appears on no price list. Net price per token is the hardest number in the model to forecast, and the one that moves the answer most.

Billable tokens multiplied by price gives the revenue the AI labs themselves earn, and the arithmetic here is still clean. 694 quadrillion billable tokens in 2030 at US$0.84 per million works out to US$583bn, up from US$55bn in 2026.

A multiplier then stacks cloud and applications on top of the labs (lever B8), set at 3.0x in the base case. That figure is not picked from the air. It is the ratio observed in 2026: US$166bn of AI revenue across the whole stack against US$55bn earned at the model layer. The base case holds today's ratio flat to 2030, which carries US$583bn of model revenue to US$1.75 trillion of total AI revenue. Holding it flat is itself the assumption. It says the industry keeps splitting its dollars between labs, clouds and software exactly as it splits them now. It also means both layers compound at the same rate by construction, so in the chart below they differ in level, never in slope.

The multiplier deserves to be named, but not oversold. Swing B8 alone from 2.2x to 4.0x, leaving every other lever at base, and 2030 revenue moves from US$1.28 trillion to US$2.33 trillion. The US$17.5 trillion of the bull case comes from moving all 43 levers at once, and mostly from price. Where B8 does decide the answer is the verdict: at 2.2x the chain still fails to cover its annual cost in 2030 (0.99x), at 4.0x it covers it comfortably (1.79x). One judgment about who captures the value flips whether this build-out ever pays for itself.

Two bookends show how wide that judgment can run. On one side, money already under contract but not yet earned: Google Cloud and AWS together hold about US$826bn of it, up 131% in two quarters. On the other, companies paid just US$12.5bn for language-model access in all of 2025, against roughly US$410bn of spending on the infrastructure. That is 33 to 1. The lab layer is solved. The layers above it are the real debate: the US$1.75 trillion of 2030 is a claim about who captures value downstream, not about how many tokens the fleet can print.

Chips carry memory by geometry, not by assumption. 13.08M packages in 2026, each holding about 255 GB of high-bandwidth memory (HBM), demand 3.34bn GB in total. That sits 11% below TrendForce's independent estimate of 3.75bn, which is the cross-check. It pulls the whole memory market from US$900bn in 2026 toward US$1.61 trillion in 2030, of which the AI share climbs from US$125bn to US$453bn.

The memory the chain drags along is no cheap accessory. DRAM contract prices multiplied 5.35x in four quarters, and Micron booked an 84.6% gross margin. That is not a typo. What drove it is a squeeze, not a fashion. HBM consumes between two and three times the silicon wafer of ordinary memory for the same gigabyte, so every stack built for AI removes capacity that would have made ordinary DRAM. Supply went short while phones, cars and ordinary servers still wanted it, and the price of everything repriced upward. The HBM market alone goes from US$42.7bn in 2026 to US$307bn in 2030. For once the bottleneck earns more than the computing that created it.

Here the consensus is exactly backwards, and the reason is the clock each market runs on. HBM sells on contracts fixed up to a year ahead; ordinary DRAM reprices every quarter. NVIDIA locked its 2026 HBM prices in May 2025, before the shock, so HBM has been less profitable than ordinary DDR5 memory since early 2026. Samsung's own results confirm it: ordinary DRAM now out-earns HBM even as HBM4 ramps for Vera Rubin. So the makers earning most from AI are not the HBM champions. They are Samsung and Micron, whose large ordinary-memory books reprice into the shock every quarter, while SK hynix, the nominal HBM leader, sits inside contracts signed before it. I think the profit in AI memory is not where the roadmaps point.

Chips draw power by spec sheet, so this step has almost no assumption in it. Only physics and the fleet from step 3. A GB200 rack pulls 132 kW across 72 GPUs, which is 1.83 kW per chip. Multiply that by each year's shipments and the capacity added follows: 21.0 GW of new computing load in 2026, on top of everything already plugged in. The additions grow every year, because the fleet grows and each new generation draws more. Racks go from 132 kW today toward 600 kW on the roadmap.

Turning gigawatts into a yearly bill takes three steps. Multiply the running gigawatts by the 8,760 hours in a year. Add 15% for cooling and losses, which is what a PUE of 1.15 means. The letters stand for power usage effectiveness: the total power the building draws, divided by the part that actually reaches the chips. Then assume the fleet is busy 71% of the time, the utilization Epoch AI measures on real fleets (lever D2). Chips idle between jobs, and assuming they run flat out would inflate the answer by 41%. The result is 207 TWh in 2026, rising to 1,238 TWh in 2030. A number that large deserves a second route, so here is one that never touches the chip count. Take the US$757bn the four largest cloud firms spend in 2026. Divide it by the roughly US$40M it costs to build a megawatt of datacenter. That gives 18.9 GW of new capacity, against the 21.0 GW the chips imply. Two independent paths, 11% apart.

That 1,238 TWh is where the physics turns into a wall. It is 3.4x the entire electricity budget the IEA sets aside for AI datacenters in 2030, and it clears the budget the agency sets for every datacenter on Earth, AI or not. Even the bear case, at 591 TWh, does not fit. The strain is already visible in hardware. A single 2027 rack will draw the peak power of 65 homes and shed the heat of 30 gas boilers. Gas-turbine orders rose 70% in 2025, with Mitsubishi sold out to 2028. And the big five now spend more on AI in 2026 than the world invests in producing oil and gas.

Two readings, one conclusion. Either the IEA's datacenter budget is wrong by a large factor, or the industry's chip growth is physically impossible. Both make the grid the binding constraint, and both hand the rent to whoever controls the connection to it. The US queue of projects waiting to plug in already exceeds 1 terawatt, and its own operators call the 'vast majority highly speculative'. In Texas alone 410 GW sit queued, 87% of it datacenters. Reports of '50% of projects cancelled' refer to exactly that unfunded queue. Measured North-American capacity forecasts moved only about 1% in six months. The ceiling is physical. And it arrives ahead of the financial one.

The public debate compares one year of spending against one year of revenue and pronounces the industry underwater. That is an accounting error. Equipment is paid for once but earns for years, so its cost is spread across those years, not charged all at once. Three different numbers get confused here, and 2026 separates them cleanly. The chain pays out US$911bn of cash that year, which is revenue for the chipmakers, the memory makers, the builders and the utilities. Spread over the years the equipment earns, that same outlay is US$270bn of annual cost. That is US$216bn of silicon over five years, US$37bn of buildings over fourteen, and US$18bn of electricity, consumed in the year it is bought. And US$166bn is what the AI layer itself sells that year.

Coverage compares the third number against the second, never against the first. It is AI revenue divided by annual cost, and it says when the chain begins to pay for itself. In 2026 it is 0.62x: US$166bn of AI revenue against US$270bn of cost, or 62 cents earned per dollar of cost. It crosses 1.0x in 2029 and reaches 1.35x in 2030, as revenue of US$1.75 trillion outruns cost of US$1.30 trillion. That crossing is the moment the arithmetic starts working in the industry's favour. Cash out the door is a separate question, and it never turns positive inside this window.

Two outside markers show how fragile that crossing is. Bain sets a stiffer bar. It puts the revenue needed by 2030 at about US$2 trillion a year, and counts US$800bn of that as still missing even after efficiency savings. Measured against Bain's bar, the base case falls 13% short. The second marker points the other way. For every US$1 of annual cost their AI infrastructure carries, the hyperscalers already book US$1.19 of revenue, above 1.00 for the first time. So the crossing is real, and narrow. It rests on the price assumption that is the least certain input in the chain, and in the bear case it never happens at all: coverage ends 2030 at 0.17x.