NVIDIA sells a whole computer room in a box, and its customers keep committing to more of them. Each new generation earns NVIDIA more money per unit of electricity, which matters because electricity is now the thing data centres run short of. The company has more cash than debt and nothing due before 2028. What has changed is that it has promised to buy, or to stand behind, far more than it has in the bank.
Source: NVIDIA Q2 FY2027 results, CFO commentary and Form 10-Q (period ended 26 Jul 2026); FY2028 guidance (26 Aug 2026); repurchase announcement (28 Sep 2026); consensus from TipRanks, MarketBeat and StockAnalysis (2–3 Oct 2026).
Why I chose it
1. It sells a rack, not a chip
A Vera Rubin NVL72 is seventy-two accelerators, a CPU complex, inference silicon, storage processors and networking, delivered as one system. The moat is fungibility: the same rack trains, serves and runs agents in any cloud, on software every practitioner already knows. Where time to deployment is the scarce input, that fungibility earns a 75.0% gross margin. Revenue per gigawatt keeps rising by generation, from about $18bn for Hopper to about $25bn for Grace Blackwell and roughly $40bn projected for Vera Rubin [industry estimate, unaudited].
In plain termsCustomers buy a finished system that works anywhere, and each new generation earns more from the same amount of power.
2. Demand is broadening beyond the four hyperscalers
AI clouds, industrial and enterprise customers reached $40.3bn in Q2, 45% of Data Center revenue and growing 138% a year, faster than the hyperscalers. Anthropic has contracted 2.5 gigawatts of NVIDIA capacity through 2028 while also buying Broadcom custom silicon and AMD accelerators. A frontier lab buying all three is a lab short of capacity, not one abandoning NVIDIA.
In plain termsThe buyers are no longer just the four biggest cloud companies, and even customers who use rival chips are still buying NVIDIA.
3. A guide that keeps being beaten
The beat has widened each quarter this year: $3.6bn above guide in Q1, $5.2bn in Q2. The Q3 guide is $108bn. Management then gave the first full-year guide in the company's history, about 70% growth for FY2028, which implies roughly $700bn of revenue against a consensus near $673bn. Shareholder returns have kept pace: the board added $150bn to the buyback on 28 September, leaving $235bn authorised through FY2028.
In plain termsThe company keeps beating its own forecasts, and has now forecast a full year ahead for the first time.
4. The credit lens: a strong balance sheet, run harder
On the snapshot, NVIDIA passes easily: about $23bn of net cash and no maturity before June 2028. On top of it sit $366bn of purchase commitments, mainly memory, and $108.5bn of guarantees, about 6.5 times liquid resources. In June it borrowed $25bn it did not need; in September it authorised $150bn of buybacks and opened talks with insurers to lay off part of its neocloud credit exposure. None of this is a solvency concern. It is why a major bank's credit desk rates the name Neutral on the day its equity desk calls it a top pick, and I think both are right.
In plain termsNVIDIA is financially strong, but it is now using that strength more aggressively. That is the part of the story most coverage skips.
What would break it
None of these threatens the company's survival. Each is observable in public data, and the most important one has a date: the 17 November cash-flow statement.
| Signal | What I would see | Where it shows first |
|---|---|---|
| Profits do not turn into cash | Operating cash flow ran at 40% of net income in Q2 and days sales outstanding rose from 45 to 60. Two consecutive quarters of cash flow below 55% of net income, or DSO above 65, would turn an earnings-quality question into a thesis problem. | Q3 FY2027 results and 10-Q, 17 Nov 2026 |
| Hyperscaler capex digestion | One large cloud company guiding 2027 capital spending flat would break the construct behind the FY2028 70% growth guide. | Alphabet, Microsoft, Meta, Amazon Q3 results, late Oct |
| Memory cost eats the margin | Management guided a gross-margin trough of 71 to 72% for Q4 FY2027 on memory cost. Micron's latest 86.8% gross margin shows who holds pricing power. A trough below 71% would mean memory was priced worse than disclosed. | Q4 FY2027 results, late Feb 2027 |
| The multiple gives way with earnings | Every scenario below holds the multiple at 20 to 21 times. Historically, when earnings disappoint the multiple is usually the first thing to give: the bear-case earnings at 15 times would be worth about $205. | Forward P/E versus its three-year average |
| Custom silicon takes share | Broadcom is targeting about $230bn of AI revenue in FY2028 from custom accelerators for the hyperscalers. Share loss would show in margin before it shows in volume. | Broadcom and Marvell results; named XPU wins |
| Circular financing | NVIDIA invests in, backstops and guarantees parts of its customer base. The loop is real but smaller than the loudest claims; the share of revenue from entities it has funded is still undisclosed. | 10-Q commitments, guarantees and related disclosures |
Known noise, not a thesis breaker: a reported US Department of Justice inquiry into the Groq licence (scope and stage unknown), and a possible Anthropic listing that would mark NVIDIA's stake to market and add volatility to GAAP earnings. Both are headline risks; neither changes the operating case on what is known today.
How I value it
FY2028 revenue about $620bn at a 72% gross margin, earnings per share near $13.70 (an 11% cut to consensus), with the multiple holding near 20 times.
The FY2028 guide is met in full: about $700bn of revenue at a 72.5% gross margin, earnings per share near $15.80, valued at about 20.6 times, just above the three-year average.
FY2028 revenue reaches about $700bn, memory cost passes through and gross margin returns to about 74%; earnings per share near $16.50 at about 21 times.
Where I sit against the street, and what the range leaves out. My base of $325 sits on the analyst mean of about $321 to $328, with no sell ratings among the majors: I think the street is right, not too high. But the range from $275 to $350 is only 32% of the share price, narrower than one year of the stock's normal movement (historical volatility near 36%), and analysts' FY2028 earnings estimates run from $9.80 to $18.75. So I also test a stress case outside the weighting, at about $155: one hyperscaler guides 2027 capital spending flat, the margin trough goes below 71%, and earnings near $11.50 are valued at 13.5 times. It is not a target; it is the outcome a reader should be prepared to survive.
Levels and dated checkpoints
The stock sits within 2% of its record, above all four rising moving averages. The levels below frame where the trend is confirmed, where it first finds support, and what would invalidate the idea on price or on fundamentals.
Glossary
- Rack / NVL72
- NVIDIA's system of seventy-two accelerators linked as one machine; the unit of sale.
- Hyperscaler
- The largest cloud operators: Amazon, Microsoft, Alphabet and Meta.
- ACIE
- AI clouds, industrial and enterprise: Data Center revenue from outside the four hyperscalers.
- Neocloud
- A specialist AI cloud operator, such as CoreWeave or Nebius, without a hyperscaler's balance sheet.
- HBM
- High-bandwidth memory, stacked on every AI accelerator; its price is the main driver of NVIDIA's guided margin dip.
- Gross margin
- The share of revenue left after the direct cost of making the product.
- Purchase commitment
- A binding promise to buy supply or capacity in future; not recorded as debt.
- Guarantee / backstop
- A promise to cover a partner's obligations or buy its unsold capacity, which helps the partner borrow.
- Circular financing
- The criticism that NVIDIA funds, lends to or guarantees the customers who then buy its chips.
- DSO
- Days sales outstanding: the average number of days between a sale and collecting the cash.
- Operating cash flow
- Cash generated by the business in a period, before investment spending.
- Forward P/E
- Share price divided by expected earnings per share over the next twelve months.
- Custom silicon / XPU
- An accelerator designed for one customer, such as Broadcom's chips for the hyperscalers, rather than a merchant GPU.
- Stress case
- A deliberately severe outcome used to test risk; not a forecast and not part of the weighted value.