Gartner sees $4 trillion of AI spending in 2027, but a positive AI return now costs 18 times more

Gartner's own release credits part of the server bill to memory-price inflation, and its analysts told chief information officers this week that positive returns now cost 18 times more to reach than two years ago. Anyone marking AI vendor revenue against the first number has to price the second.

Richard TangRichard Tangthree-minute read
Share
Gartner Distinguished VP Analyst Mary Mesaglio speaking on stage at Gartner IT Symposium/Xpo, gesturing mid-sentence in front of a dark blue backdrop
1 / 7Slide 1 of 7
Gartner Distinguished VP Analyst Mary Mesaglio presenting at Gartner IT Symposium/Xpo in Orlando in October, where the firm's analysts walked chief information officers through their AI findings. This week's Kochi symposium put a number on the other side of the bill: 18 times more spending to reach positive AI returns than two years ago.

Within five days, Gartner−2.16% — Gartner, down 2.16 percent as of its last quote published both halves of the AI trade's ledger. On November 12 the research firm forecast $4.05 trillion of global AI spending for 2027, 59% of it infrastructure 1; today, at its IT Symposium/Xpo in Kochi, India, its analysts reported that companies now spend 18 times more to reach positive returns on AI than two years ago 2.

The forecast is the demand anchor the AI supply chain trades on; the symposium findings say what that spend is actually buying. Coming from the same firm days apart, each softens the other.

Infrastructure takes 59% of Gartner's $4.05 trillion 2027 AI forecast

Infrastructure alone outspends every other segment combined

  • Estimate
$0B$500B$1,000B$1,500B$2,000B$2,500BAI infrastructure$2,397BAI services$746BAI software$649BAI cybersecurity$86BAI agents and assistants$67BGenerative AI models$51B
Data
Value
AI infrastructure (estimate)$2,397B
AI services (estimate)$746B
AI software (estimate)$649B
AI cybersecurity (estimate)$86B
AI agents and assistants (estimate)$67B
Generative AI models (estimate)$51B
Gartner worldwide AI spending forecast for 2027, in billions of US dollars, published November 12, 2026. The bars are the six largest segments Gartner lists; smaller categories account for the remaining $54 billion of the $4.05 trillion total.1

The server bill is partly a memory-price story

The infrastructure line is not only more machines. Spending on AI-optimized servers climbs from $528 billion this year to $957 billion in 2027, and Gartner credits part of the rise to what it calls memflation, rising memory prices lifting server prices 13. The firm said in April that relief was unlikely before late 2027, and its chief of research, John-David Lovelock, said in October the squeeze could persist into 2030 3. The release never splits price from volume, so the inflation slice cannot be sized from it 1.

The forecast touches the server suppliers Nvidia−3.99% — Nvidia, down 3.99 percent today, Dell and Super Micro−6.80% — Super Micro, down 6.80 percent today 3, and part of what reads as demand growth on that line is price inflation.

Nvidia and Super Micro revenue has multiplied fivefold since late 2023

  • Nvidia
  • Dell
  • Super Micro
$0B$50B$100BQ3 '23Q1 '24Q3 '24Q1 '25Q3 '25Q1 '26Q3 '26$96.22B$46.97B$19.95B
Data
NvidiaDellSuper Micro
Q3 '23$18.12B$22.25B—
Q4 '23$22.1B$22.32B$3.67B
Q1 '24$26.04B$22.24B$3.85B
Q2 '24$30.04B$25.03B$5.36B
Q3 '24$35.08B$24.37B$5.94B
Q4 '24$39.33B$23.93B$5.68B
Q1 '25$44.06B$23.38B$4.6B
Q2 '25$46.74B$29.78B$5.76B
Q3 '25$57.01B$27.01B$5.02B
Q4 '25$68.13B$33.38B$12.68B
Q1 '26$81.62B$43.84B$10.24B
Q2 '26$96.22B$46.97B$11.12B
Q3 '26——$19.95B
Quarterly revenue in billions of US dollars for the three server suppliers Gartner's forecast names, by calendar quarter of the period reported. Nvidia and Dell report fiscal quarters ending within weeks of the calendar quarter and had not reported the September 2026 quarter when this was published; Super Micro's fiscal year ends in June, so its latest period is the September 2026 quarter. Sources: Nvidia, Dell Technologies and Super Micro filings with the SEC, retrieved November 16, 2026.8,9,10

Buyers still convert one project in five

On the paying side, conversion has not moved. Neha Kumar, Senior Director Analyst at Gartner, told the Kochi audience the success rate is still one project in five, unchanged from the same stage a year ago, and only 52% of more than 1,000 surveyed leaders report any positive value at all 2. That sits in the same family as the finding we reported on October 29, that 97% of companies could not turn AI into savings 4, so the stall reads as a trend rather than a one-off.

The buildout needs $8.5 trillion it will not earn

Gartner's October analysis already priced the supply-side consequence: AI data center owners need $8.5 trillion of revenue through 2030 to cover announced capital spending, against at most $3.6 trillion it expects them to earn, and it expects more than 80% of neoclouds, the specialized AI cloud providers, to stop operating independently by 2030 5. The missing link is the redesign of work between buying compute and banking savings; as Gartner VP Analyst Nate Suda put it in Kochi, time saved is not money saved 2.

The tollbooth collects on both sides

Gartner sells research to the vendors who quote the forecast and to the chief information officers who must convert it, so both camps keep paying it. Its shares closed the week at $192.15 on a hold consensus, with full-year results due February 2 6; the subscription book we examined on November 3, contract value up 2.5%, is the quiet core underneath 7. This week gave both camps a fresh reason to renew.

Underwrite AI infrastructure at the $2.4 trillion line only after the two haircuts its own source applied in the past five days: memory-price inflation inside the bill, and conversion failure behind it 12. The forecast already assumes the spend, so what actually moves vendor revenue is whether that one-in-five success rate changes 2. The buildout is still worth backing, because capability keeps compounding, and buyers who redesign work will keep the gains that Gartner warns could otherwise leak to vendors and outsourcers 2.

Get Data Center by email, free

Who is building AI's data centers, and who is paying for them. In our daily email, every weekday.

Loading IT

Reader comments

0 comments

    More in Data Center

    All in this topic →

    Data Center

    MARA sold 13,597 bitcoin for $1.1 billion, cut debt 37%; two AI leases by Dec. 31 decide the stock

    Traders blamed the miner for bitcoin's crash to $62,006, but its $1.1 billion coin sale is a balance-sheet story: lenders repaid in full, a bitcoin treasury converting into AI data centers, and a December 31 lease deadline that decides the stock.

    Data Center

    Rackspace signed a $380 million Together AI deal to buy Nvidia GB300 racks; no lender has committed

    The definitive agreement looks like a win for Rackspace, but the company said the GB300 hardware will be financed separately and no lender commitment exists yet. Until that loan closes, the $610 million AI book and the megawatt target behind it are conditions to be met, not numbers to book.

    Data Center

    Everpure CEO and co-founder trusts sold $17 million eight days before Wednesday's earnings report

    Forms filed November 12 show $17.1 million of insider sales on November 10, at $98.88 to $104.75, about a third below the October peak and eight days before the November 18 report. The report has to reconcile a 50 percent growth call from Guggenheim with a company that burned $136 million of operating cash last quarter.

    Data Center

    Ciena splits its product-and-technology job 24 days before year-end results

    Greg Friesen becomes chief product officer while Brodie Gage exits, and no one is named over technology as an $8.5 billion backlog waits on components that stay short into 2028. The split lands 24 days before the results that have to turn that backlog into revenue.

    Data Center

    Goldman Sachs cut IREN's target to $30 and disclosed a 7.2% stake in the same week

    The bank that repriced the neocloud sector's earnings also reports holding 28.4 million IREN shares, in the ordinary course, with no plan to buy or sell. IREN holders now weigh a $20.3 billion backlog against $6 billion of unfunded build.

    Data Center

    Jabil and Adani sign a 50:50 India AI-hardware MoU as data centers hit 49.7% of revenue

    Data-center hardware is now 49.7% of Jabil's revenue, but it assembles at a 5.1% trailing operating margin while component makers like Legrand earn 20.8% on the same layer. The December quarter, guided to $10.6-11.4 billion, is the first test of the mix.

    Subscribe to TheInference for free