How Big Is the AI Server Market? Public Answers Span $35bn to $223bn

How big is the AI server market in 2026? Put the publicly available institutional answers side by side and they run from roughly $35bn to roughly $223bn — the same question, a factor of six apart.

That is not a research-quality problem. "AI server" simply has no agreed definition: does it include networking? Full liquid-cooled racks? Chip factory-gate prices or system street prices? Every institution is honestly answering the question as it defined it — and what travels is just the number.

Why this deserves attention

Because "the market is $X bn" is the foundation under a large share of investment narratives. Derive a company's share, penetration or ceiling from it and you silently inherit an arbitrary definition — with $35bn or $223bn as the denominator, the same company's "penetration" differs six-fold.

Worse, these numbers detach from their definitions as they travel. The original report usually states what is in and out; three citations later, all that survives is "analysts see the AI server market reaching $X bn."

The alternative: cite nobody — build the number up from filings

Three US filers happen to disclose an AI revenue line quarterly, each on an audited or filed basis:

• NVIDIA's "Data Center" revenue (quarterly 8-Ks) — note it includes networking;
• AMD's Data Center reporting segment (segment tables in each release, to the million) — note it includes server CPUs, wider than pure AI;
• Broadcom's "AI semiconductor" revenue (company-defined, disclosed on calls and in releases) — custom accelerators plus AI networking.

The three lines overlap imperfectly (NVIDIA and Broadcom both include networking), so their sum is an order of magnitude, not a precise TAM. But it has one property no institutional estimate can offer: every input can be re-verified in SEC filings, quarter by quarter.

A worked example: splicing NVIDIA's data-center revenue into calendar years

Take the largest line. NVIDIA's fiscal quarters end in late January, April, July and October, so quoting "FY2025 data-center revenue" misaligns with everyone else's calendar year by a quarter. The fix: map each fiscal quarter to the calendar quarter covering most of its months, then add them up:

• Calendar 2023: $47.5bn
• Calendar 2024: $115.2bn
• Calendar 2025: $193.7bn

All three numbers are sums of publicly filed quarterly figures; anyone can reproduce them in half an hour. Four-fold in two years — that is the filing-grade version of "the AI hardware market is exploding," with no institution cited.

Three traps when reproducing this. One: fiscal offsets differ by company (Broadcom's quarters end in early February, May, August, November), so write the mapping rule per company. Two: never add differently-defined lines into a "precise total" — the overlaps make every decimal place fake; "order of magnitude" is the honest unit. Three: the sum is a floor — Google's TPUs and Amazon's Trainium have no revenue filings and are invisible to any bottom-up build.

What the method can and cannot carry

Can: the direction and slope of growth (filed quarterly, beyond dispute); a floor on market magnitude; a sanity check on whether any institutional number is obviously off.

Cannot: a precise TAM (overlaps plus invisible in-house silicon); cross-layer comparisons (memory, systems and networking each keep separate books); anything that needs share measured to the percentage point.

Methodologically this is one discipline applied once more: result-layer numbers (institutional estimates) serve as an envelope; the load-bearing goes to cause and transmission layers (filings). The institutional numbers are not wrong — quoting them as definition-free facts is.

Where this sits in the report

The arithmetic above is the methodological skeleton of chapter seven of our Global AI Compute Value Chain deep dive (second edition). The report ships as two PDFs (30 pages in Chinese, 34 in English; 19 chapters + 4 appendices, 13 figures and 14 data tables), reference date the 27 August 2026 close; the ten chokepoint companies' full quarterly series, the complete bottom-up market build and the six-fold-split chart itself are inside, with every self-computed figure derived step by step in an appendix.

To see how we grade evidence and run the process first, start with our research method. Related method pieces: when a key metric vanishes from the earnings release and the blank field in a Form 144 that decides everything.

Data and sources. NVIDIA data-center revenue from its quarterly 8-Ks and CFO Commentaries (2023–2026, SEC EDGAR); AMD segment data from quarterly release segment tables; Broadcom AI revenue from its releases and calls. The calendar-splice mapping is this article's method and is reproducible from the same public documents. The institutional range is an observation across public research; this article adopts none of its points. Accessed 28 August 2026.

Disclaimer. This is research content. It does not constitute investment advice, an offer or a solicitation, and is not tailored to any particular investor's financial situation, objectives or risk tolerance. We hold no investment-adviser licence in any jurisdiction. Past performance does not indicate future results.