Key takeaways
- Epoch AI puts GPT-4's final training run at $40 million. Microsoft's fiscal 2026 additions to property and equipment were $115.9 billion, nearly 2,900 times as much.
- No 10-K or 10-Q splits AI spending between training and inference. Microsoft's filing names 'AI training and other infrastructure' without a dollar figure on either side.
- Alphabet, Meta and Amazon disclosed a combined $227.8 billion of capital expenditure in the first half of 2026, and Meta guides to $130 billion to $145 billion for the year.
- The spending reaches profits as depreciation: Microsoft charged $34.3 billion in fiscal 2026 against $15.2 billion two years earlier, and Microsoft Cloud gross margin slipped to 66%.
- On Epoch AI's 2016 to 2024 sample, frontier training costs grew 2.4x a year, with the largest single runs projected to pass $1 billion by 2027.
The honest answer: the split isn't disclosed, but the filings bound it
Which costs more, training an AI model or running it for users? That question sits underneath every AI capex headline, and no company that knows the answer has published it. What's on the record instead is a set of totals. Microsoft's latest 10-K shows $115.9 billion of additions to property and equipment for the fiscal year ended June 2026, up from $64.6 billion the year before. The best training-side numbers, meanwhile, are academic estimates: Epoch AI's 2024 cost study puts the final training run of GPT-4 at roughly $40 million.
Set those two figures side by side and the shape of an answer appears. Record training runs are measured in tens of millions of dollars. Infrastructure budgets are measured in hundreds of billions. Training alone can't absorb the money these companies are spending, so most of it is going somewhere else: inference, meaning the work of serving a trained model's answers to users, plus research experiments and capacity built ahead of demand. In what proportions, nobody outside those companies knows. This piece reconciles the two kinds of number, the disclosed and the estimated, and shows where each one runs out.
The AI training vs inference cost split that no filing provides
Start with what the documents actually say. Microsoft's fiscal 2026 annual report comes closest to naming both workloads, committing to "capital expenditures to support growth in our cloud offerings and our investments in AI training and other infrastructure." Training is named. Everything else is "other". Neither side carries a number.
Amazon's 10-Q for the June 2026 quarter attributes its capex mostly to technology infrastructure, "the majority of which is to support AWS business growth", with no workload split. Alphabet describes purchases of "land, buildings, and servers and network equipment". Meta's guidance covers "our AI efforts and core business" in a single phrase. NVIDIA sells the hardware both workloads run on, and its most recent 10-Q, for the quarter ended April 2026, doesn't contain the word "inference" at all; "training" appears twice, both times in risk-factor language about restrictions on where its products can be sold and used.
That's the whole public record. A reader of the primary documents gets totals and adjectives, not an AI training vs inference cost breakdown. The totals still earn a careful read, because they put bounds on what each side could plausibly be.
What the buyers disclosed: $227.8 billion of capex in six months
Four companies' 2026 filings carry most of the AI infrastructure spending story. Microsoft's additions to property and equipment reached $115.9 billion in fiscal 2026, against $64.6 billion in fiscal 2025 and $44.5 billion in fiscal 2024, a 2.6x increase in two years. Alphabet spent $80.6 billion on capital expenditure in the six months to June 2026, more than double the $39.6 billion of the same period a year earlier. Amazon reported $96.3 billion of cash capex for those six months, up 73% from $55.6 billion. Meta spent $50.92 billion in the half and anticipates approximately $130 billion to $145 billion for 2026 as a whole.
Add up the three companies that report on calendar quarters and you get $227.8 billion of disclosed capital expenditure in six months. The chart above shows the four disclosed totals side by side. Set Epoch AI's costliest training-run estimate against any of those bars and it disappears: $40 million is less than 0.04% of Microsoft's bar alone.
Two further disclosures point forward. Microsoft's 10-K lists $329.1 billion of leases, primarily for datacenters, signed but not yet commenced as of June 2026, with commencement dates running out to fiscal 2033. And in June 2026, Alphabet raised $49.6 billion of new equity for purposes including "capital expenditures to scale AI infrastructure and global compute". A company with Alphabet's cash flows selling stock to fund servers is itself a data point, and the strain surfaces in Google Cloud operating margin before it surfaces anywhere else.
What published estimates say a frontier training run costs
The training side has real numbers, but they're estimates rather than disclosures. The most cited estimate of the cost of training a frontier AI model comes from Epoch AI's 2024 study, revised in February 2025, which models hardware, energy, cloud rental and staff costs. Its central finding: the most expensive publicly announced training runs to date were GPT-4, published in March 2023, at roughly $40 million, and Google's Gemini Ultra at roughly $30 million. Across its 2016 to 2024 sample, the amortised cost of the most compute-intensive runs grew at 2.4x per year, with a 90% confidence interval of 2.0x to 2.9x. Projected forward, the largest single runs pass $1 billion by 2027.
The composition matters as much as the total. Epoch finds that accelerator chips and research staff are the two largest costs, each running to tens of millions of dollars for a frontier model. Server components add 15% to 22% of the total, cluster-level interconnect 9% to 13%, and energy just 2% to 6%. Widen the lens from the final run to the whole development effort, experiments included, and staff costs including equity make up 29% to 49% of the amortised total. Training a frontier model is as much a payroll problem as a hardware problem.
Now reconcile. Microsoft's fiscal 2026 capex could fund nearly 2,900 runs at GPT-4's estimated cost. Even a $1 billion run of the kind Epoch projects for 2027 is under 1% of what Microsoft alone spent in fiscal 2026. Whatever these companies bought with a quarter of a trillion dollars in six months, it wasn't mostly final training runs.
Depreciation and margins are where the bill actually lands
Capex is a cash-flow item; the income statement meets it later, as depreciation spread over each asset's assumed life. That's where the AI bill has started to arrive. Microsoft's depreciation expense was $34.3 billion in fiscal 2026, from $22.0 billion in fiscal 2025 and $15.2 billion in fiscal 2024. Alphabet charged $13.6 billion of depreciation on property and equipment in the first half of 2026, up 43% year on year. Depreciation on servers is the fastest-moving piece: Meta's servers and network assets alone carried $9.01 billion of it in the half, up 57% from $5.74 billion.
The assumed lives are short and contested. Microsoft depreciates servers and network equipment over two to six years. Meta extended the useful lives of most of its servers and network assets to 5.5 years from January 2025, and that single change of estimate cut its 2025 depreciation by $2.92 billion and added $2.59 billion to net income. The sensitivity runs both ways, which is why AI capex depreciation assumptions deserve a closer read than the capex headlines themselves.
The margin consequences are already visible. Microsoft reports that its Microsoft Cloud gross margin percentage decreased to 66% in fiscal 2026, "driven by continued investments in AI infrastructure and growing AI product usage". The company selling the hardware shows the mirror image: NVIDIA's gross margin was 74.9% in its quarter ended April 2026, on Data Center revenue of $75.2 billion, up 92% in a year, with roughly half of it from hyperscalers. That's about $37.6 billion of a single quarter's chip revenue coming from a handful of cloud buyers. A broad index holding carries exposure to both sides of that $37.6 billion transfer: NVIDIA's 74.9% margin on the sale sits in the same portfolio as the buyers' years of depreciation on the purchase. How the transfer nets out is the story told by the AI gross margin that software companies report.
The strongest counter-case: the ledger can't see workloads
There's a serious objection to this whole framing, and the filings themselves supply it. Company accounts don't meter GPU hours; they classify spending by function. Microsoft's operating expenses rose $4.9 billion in fiscal 2026 partly on "continued investments in research and development compute capacity", which is where training experiments naturally sit, while the cost of serving customers lands in cost of revenue. Alphabet's R&D expenses grew $7.9 billion in the first half of 2026, and $1.1 billion of that growth was depreciation. The same server shows up in different lines depending on who books it and what it ran last month. Nothing in the accounts stops a cluster bought for training from serving users afterwards; the capex was one purchase either way.
Epoch's numbers carry the matching caveat from the other direction. Its headline figures price the final training run. The study's own wider accounting finds the full development effort, failed experiments and staff included, is a multiple of that, with staff at 29% to 49% of the amortised total. So "training cost" isn't one number. The run, the development effort and the depreciating cluster give three different answers, and none of them is a line item any filing reports. The AI training vs inference cost question may simply be badly posed at company level. The honest version is narrower: how much of the installed base earns revenue today, and how much is a bet on future workloads of either kind?
What these numbers cannot tell you
The reconciliation above leans on estimates exactly where it matters most. Epoch's figures are model-based, published in 2024 for models released in 2023, and the paper's own confidence interval on cost growth runs from 2.0x to 2.9x a year. It says nothing about the runs of 2025 and 2026, which its own trend puts several times higher. The capex totals aren't all AI, either. Amazon's figure explicitly includes its fulfilment network. Alphabet's includes land and buildings. Meta's range is guidance, not outturn.
Fiscal calendars don't line up: Microsoft's year ends in June, NVIDIA's quarter ended in late April, and the rest report calendar quarters, so every comparison here is between overlapping rather than identical periods. Depreciation depends on useful-life assumptions that companies revise, as Meta's 2025 change shows. And capex plans, margins and guidance are all statements by interested parties. The one number nobody audits is the one this piece is about, because no auditor is ever handed it.
What would change the conclusion
If Epoch's 2.4x annual growth in training costs holds, the arithmetic flips within a few model generations. A $1 billion run in 2027, compounding at that rate, needs only a few more doublings to reach the tens of billions, at which point training stops being a rounding error against even Microsoft's capex.
If any filer starts splitting depreciation or revenue by workload, the estimates here become measurements overnight. That requires no new science, only a new disclosure line, and quarterly reports have absorbed new lines before.
If useful lives shorten, the margin arithmetic moves fast. Meta's move to a 5.5-year server assumption added $2.92 billion back to one year's pre-tax profits; the same lever, pulled the other way by faster chip cycles, subtracts at the same scale.
The near-term tell sits in two numbers in the next few filings: Microsoft Cloud gross margin, at 66% in fiscal 2026 with AI investment named as the drag, and the pace at which Microsoft's $329.1 billion of not-yet-commenced leases converts into depreciation. The first measures whether serving models earns its keep; the second measures how much more of the bill is already signed for. When those two lines stop moving against each other, the AI training vs inference cost question has its answer, in the only place it was ever going to appear: the filings.