NVIDIA's datacentre revenue falls year on year

Where things stand (as of 12 Oct 2026)

• Latest print. In Q2 FY27 (quarter ended 26 Jul 2026), Data Center revenue was $89.0B, up 117% year over year ($89,023M against a recast $41,096M). NVIDIA now reports two platforms, Data Center (split into Hyperscale and ACIE) and Edge Computing. It adopted this in Q1 FY27 and recast prior periods. Source: the NVIDIA Q2 FY27 release (nvidianews.nvidia.com, 26 Aug 2026) and the CFO-commentary split (gpuinsights.net, 2 Oct 2026).
• Guidance. Q3 FY27 revenue is guided at $108B ±2%, which implies roughly $100B of Data Center revenue. On the call the CFO said NVIDIA expects about 70% revenue growth in FY28 (calendar 2027). She called it "supply-constrained": customer forecasts point to roughly double that, and supply stays tight at least through the end of FY28 (CNBC, Fortune, Axios, 26 Aug 2026; thesoftwarefrontier.com, 20 Sep 2026).
• Analyst consensus. Revenue goes from $412B in FY27 to $1.21T in FY31, i.e. no decline is forecast (Motley Fool via Yahoo, about 5 Oct 2026).
• Year-ago comparators for the next reports, from the OWID series of NVIDIA filings:
  • Q3 FY26: $51.2B
  • Q4 FY26: $62.3B
  • Q1 FY27: $75.2B
  • Q2 FY27: $89.0B
• Calendar. There is one report per calendar quarter. The first eligible quarter is Q3 FY27 (ends 25 Oct 2026; results about 17 Nov 2026).
• 53-week year. The 10-Q confirms FY27 is a 53-week year with a 14-week Q4. So Q4 FY28 (13 weeks, reported Feb 2028) faces an inflated comparison. That only matters if a slowdown happens to line up with it.

What that means near term: before mid-2028, YES would need Data Center revenue to fall 25–40% from its peak while NVIDIA is supply-constrained. Only a shock could do that: a Taiwan blockade, a credit crisis, or the collapse of a major AI lab.

Base rate

I downloaded OWID's NVIDIA quarterly segment series (https://ourworldindata.org/grapher/nvidia-quarterly-revenue-segment.csv), which is built from NVIDIA filings, and counted year-over-year changes.

• Data Center: 5 of 46 quarters were down year over year, in two episodes (calendar 2015 and 2019). That is about one episode every 5.75 years.
• NVIDIA total revenue: also two episodes (2019 and 2022–23).
• Semiconductor industry: downturns come roughly every 4 years.
• Hardware businesses growing more than 50% a year (Cisco 2000, NVIDIA Data Center 2018, Micron 2018, AMD 2022, Super Micro 2024) usually posted a year-over-year decline within five years. The exception is NVIDIA's own Data Center business after 2020, which was rescued by generative AI.

Treating episodes as random events at about 0.17 a year gives roughly 58% over five years. Knocking out the next ~1.5–2 years, which are close to riskless given guidance, leaves about 45%.

Paths to YES

1. Spending plateaus in 2028–30. Morgan Stanley (28 Sep 2026) projects hyperscaler plus neocloud capex growth of 93% in 2026, 61% in 2027 and 14% in 2028. After a 70% FY28 ramp, the exit rate sits far above the annual average. A FY29 with only mid-teens growth would therefore be flat quarter to quarter, and Q4 FY29 or the FY30 quarters could come in below the prior year. The question counts any decline, however small.
2. A financing-driven bust. There are late-cycle signs:
   • NVIDIA has $99B of equity stakes in AI companies, take-or-pay and revenue-share deals with AI clouds, and guarantees of up to $108.5B (including $105B for an OpenAI-leased campus in Ohio).
   • Its days sales outstanding rose from 45 to 60 because large customers were given longer to pay.
   • The 10-year Treasury yield is at a 24-year high of about 5.3%, the Fed hiked in September, and the Iran war has pushed oil up.
   • OpenAI's annualised revenue came in at about $50B against roughly $70B the market expected (FT/Yahoo, 8–9 Oct 2026). Oracle sent a force-majeure notice on its New Mexico project and its $18B of loans there trade at about 90 cents on the dollar (Reuters, 18 Sep 2026).
   • Polymarket gives an "AI bubble burst" (a defined stock crash) by 30 Jun 2027 an 18.5% chance.
3. Share loss to TPUs, Trainium, AMD and Broadcom custom chips while overall spending stalls.
4. Shocks: a Taiwan disruption, or regulation driven by AI-safety concerns.

Paths to NO

• AI revenue is compounding: Anthropic was reportedly at a ~$65B run-rate and OpenAI at ~$50B.
• NVIDIA's revenue per gigawatt keeps rising: about $18B for Hopper, $25B for Blackwell and $40B for Vera Rubin.
• Power and memory shortages stretch out the ramp rather than ending it.
• Customers are broadening beyond the hyperscalers.
• NVIDIA has grown Data Center revenue year over year for 26 straight quarters, and consensus shows growth through FY31.

Forecast construction

I modelled a probability for each quarter that the first decline appears then:

Period	Hazard per quarter
2026–27	0.1–0.6%
2028	rising from 0.9% to 3.3%
2029	about 5.5% (where a plateau after the FY28 ramp would first show up)
2030	about 5%
2031	about 4.5%

Accumulating these gives cumulative probabilities of about 1.6% by end-2027, 9% by end-2028, 28% by end-2029, 41% by end-2030 and 51% by end-2031.

Strongest case against

For lower odds: NVIDIA keeps beating its own forecasts. If AI turns out to be a decade-long supercycle, with spending reaching the $3–4T a year by 2030 that Huang talks about, there may be no decline before the 2030s. That would point to about 30%.

For higher odds: industry base rates and the late-cycle signals argue for 60–70%.

I weight these roughly equally, which lands near even odds.

Question. By the last day of each quarter, has NVIDIA reported a fiscal quarter in which Data Center revenue was lower than in the same fiscal quarter a year earlier?

Definitions.

• Data Center revenue: NVIDIA Corporation’s worldwide revenue for its reported Data Center market platform, including both compute and networking, in nominal US dollars. This is not total corporate revenue, the broader Compute & Networking segment, or revenue from a single product or geography. A renamed category counts if NVIDIA identifies it as the successor to Data Center; split categories may be summed only if NVIDIA identifies them as collectively covering that category without overlap.
• Fiscal quarter: one of NVIDIA’s reported Q1–Q4 periods, compared with the same numbered quarter of the preceding fiscal year, without adjusting for differences in weeks. The Observation period convention applies.
• Lower: a strictly negative year-over-year change, with no minimum percentage decline. Use the most precise published revenue figures; an explicit statement of a year-over-year decline also suffices where rounding makes the displayed amounts equal.

Resolves YES if NVIDIA publishes actual quarterly results meeting that comparison. Use the prior-year comparator supplied in the same report, including any recast comparator; if none is supplied, use NVIDIA’s latest published figure for that prior-year quarter on the same reporting basis. A later official revision can first establish YES for an eligible observation period. Subsequent revisions do not reverse a YES under the Resolution window convention. No attribution to reduced AI demand is required.

Resolves NO if no qualifying report has been published by the horizon. Guidance, analyst estimates, sequential declines, slower but still positive year-over-year growth, equal revenues, declines confined to a subcategory, and annual or year-to-date declines without an identifiable quarterly decline do not count. Historical quarters ending on or before the opening date do not count, even if republished or revised later. If a reporting change prevents a like-for-like category comparison and NVIDIA supplies no comparable figures, that report does not establish YES.

Resolving source. NVIDIA-authored quarterly or annual financial-results releases and accompanying financial schedules or CFO commentary on nvidianews.nvidia.com or investor.nvidia.com, and NVIDIA’s Forms 10-Q, 10-K, or results exhibits to Form 8-K filed with the US SEC (CIK 1045810). The Data Center highlights in NVIDIA Announces Financial Results for Second Quarter Fiscal 2027 identify the relevant series. These NVIDIA-authored sources alone qualify under Public confirmation (A2); third-party reports do not substitute. Apply Resolution date (A3), the Resolution window convention, and the Observation period convention.