Thesis. Compute is revenue and the P&L proves it — a $96.2B quarter, +106%, a fourth straight quarter of accelerating growth, and FY28 guided at ~70% on a supply-constrained basis. The only company that sits above every bottleneck on this watchlist.
Reviewed 19 Sep 2026, after Q2 FY2027 results on 26 Aug. Next review after Q3 FY2027 earnings on 18 Nov 2026.
What it does
NVIDIA is the platform of the AI era, not merely the best chip: GPUs, networking, CUDA software and rack-scale systems sold as one stack that the entire industry builds on. Fabless — TSMC manufactures everything. CEO Jensen Huang.
CUDA is twenty years of compounded software: cuDNN, TensorRT-LLM, NCCL and the libraries every training run depends on. It is the reason a competitor with comparable silicon still struggles to take share in training.
Two segment names matter when reading the prints. Hyperscale is the big four clouds. ACIE — AI clouds, industrial and enterprise — is everything else, and it now includes the neoclouds in this file. ACIE is growing faster than hyperscale, which is a genuine change in who is buying.
The debate is no longer whether the business is good. It is whether custom ASICs, the China loss, or the law of large numbers can dent it — and whether $4.7T is already the reward.
Bull case
- Q2 FY2027 revenue $96.2B, +106% YoY and +18% QoQ, beating $92.17B consensus by ~$4B. Data Center $89.0B, +117% — 92% of the company
- Gross margin 75.0% on both GAAP and non-GAAP. Non-GAAP EPS $2.22 against $2.10 expected; GAAP EPS $2.46, +128%
- Q3 guided $108B ±2% at 74.0% gross margin — above consensus, and assuming zero Data Center compute revenue from China
- Fourth consecutive quarter of accelerating year-over-year growth, and a fourteenth consecutive quarter above its own outlook
- FY2028 revenue growth guided at ~70%, explicitly described as supply-constrained — the constraint is capacity, not demand
- AWS will buy 2 million GPUs and adopt the new Vera CPU, with a further 2M deploying through Q2 FY29. Data Center CPU revenue expected to more than double in FY28
- ACIE $40.0B, +25% sequentially and +138% YoY as neocloud partners bring capacity online — now ~45% of Data Center and growing faster than hyperscale ($49.0B, +13% seq)
- Vera Rubin ramping; NVL72 validated at CRWV. Cadence intact
- Returned a record $26B to shareholders with ~$99B of authorisation remaining. Fortress net cash
Bear case
- ~40% of revenue comes from four hyperscalers who are all building their own chips through AVGO and MRVL. Share is drifting from ~80–90% toward ~75%, and every custom accelerator programme is revenue NVIDIA does not get
- Gross margin is guided down. CFO Colette Kress expects it to bottom in Q4 FY27 at 71–72%, explicitly on memory prices: "Memory scarcity today is being driven in large part by the AI buildout itself. The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year."
- China is at zero, having been 20–25% of datacentre revenue. Huawei Ascend is ramping into that gap
- CUDA's moat is narrowing for inference — Triton, torch.compile and ROCm are within 10–30%. Training remains defended
- A $4.7T market cap brings the law of large numbers, and this is the most-owned stock in the market, which makes it sentiment-sensitive in both directions
- The supply-constrained framing cuts both ways: it flatters demand today and means any capacity slip is a revenue miss
Major customers
Hyperscalers — the bulk of datacentre revenue:
- Microsoft, Meta, Amazon, Google, Oracle. Several have individually exceeded 10% of revenue. ⚠️ NVIDIA discloses concentration in aggregate — verify named percentages in the 10-K
Neoclouds and AI specialists:
- CRWV, IREN, NBIS and other GPU cloud operators
- OpenAI, xAI, Anthropic and other labs, directly and through cloud partners
OEM channel: Dell, HPE, Supermicro, Lenovo.
Read-through — NVIDIA is the hub of this entire file. It is a customer of TSM (manufacturing), MU and SKHY (HBM), COHR and LITE (optics — it invested ~$4B across both to secure supply), GLW (fibre, 10x capacity expansion) and VRT (rack power and cooling). It is a supplier to the neoclouds. It competes with AVGO and MRVL in custom silicon and AMD in accelerators.
The single most valuable disclosure for this file is the memory comment. The most powerful buyer in the supply chain is saying it cannot avoid paying up for memory — direct demand-side confirmation of what SNDK, MU and SKHY report from the supply side. The memory makers are extracting price from NVIDIA itself.
Practical use: NVIDIA's guidance is the demand input for nearly every category here. Treat it as a bellwether, not a category constituent.
What would change the view
- Hyperscaler capex and in-sourcing pace — how fast custom ASIC programmes actually displace merchant GPUs
- Gross margin against the guided 71–72% Q4 FY27 trough, and whether memory costs keep rising past that
- Vera Rubin timing and ramp
- Networking attach rate, the quiet part of the full-stack story
- China policy, though the outlook already assumes zero
- Whether ACIE keeps outgrowing hyperscale — the neoclouds becoming the marginal buyer is a structural shift worth tracking
- Any deceleration in the sequence of accelerating growth quarters
Update log
19 Sep 2026 — Converted to the standard template. Removed the superseded Q1 FY2027 snapshot and thesis-tracker table, which were still quoting $81.6B revenue and a $91B guide against the current $96.2B actual and $108B guide.
7 Sep 2026 — Verdict unchanged: Conviction. The memory-cost guidance identified as the most important cross-read in the file.
26 Aug 2026 — Q2 FY2027 reported. Revenue $96.2B (+106%), Data Center $89.0B (+117%), gross margin 75.0%, non-GAAP EPS $2.22 (vs $2.10), GAAP EPS $2.46 (+128%). Q3 guided $108B ±2% at 74.0% GM, assuming zero China. FY2028 growth guided ~70%, supply-constrained. AWS committing to 2M GPUs plus Vera CPU, with 2M more through Q2 FY29. Hyperscale $49.0B (+13% seq); ACIE $40.0B (+25% seq, +138% YoY). Record $26B returned; ~$99B buyback authorisation remaining. Gross margin guided to bottom at 71–72% in Q4 FY27 on memory prices.
Probabilities: strengthen ~60% / weaken ~35%.
Research and education only — not investment advice.
Made with Bullet