Nvidia Stock Nears $6 Trillion as Rubin Targets Cheaper Inference

A rally that history says may not be the full story

In late September, the main complaint about Nvidia was that its stock had gained only about 20% in 2026. That was still ahead of the S&P 500 and the Nasdaq Composite, but it was modest for a company investors expect a lot from.7 Two weeks later the picture had changed. Nvidia closed at a record $238.90 on October 5, which valued the company at roughly $5.8 trillion, about 4% below a $6 trillion mark that no company has ever reached.4 Depending on the day and the outlet, the year-to-date gain now sits between 27% and nearly 30%.169

The question is what happens next. The historical record points to one clear answer: the stock price has followed earnings and valuation, and milestones have meant little. A Motley Fool review covered the five times a U.S. company became the first to reach a new trillion-dollar level: Apple at $1, $2 and $3 trillion, and Nvidia at $4 and $5 trillion. Twelve-month returns after those milestones ranged from a 31% loss to a 27% gain, and the stock beat the S&P 500 only twice.4 In all five cases, the price-to-earnings ratio was lower a year later. What separated good years from bad ones was whether earnings grew fast enough to offset that shrinking multiple.4

That framing fits what has happened in 2026. Nvidia's business has grown much faster than its share price, and the gap between the two is now the core of the bull case.

The numbers behind the compression

Revenue for the fiscal second quarter, which ended July 26, rose 106% from a year earlier to $96.2 billion. That was faster than the 85% growth of the quarter before.2 Adjusted earnings per share rose 120%, to $2.22.4 Management has guided third-quarter revenue to about $108 billion, give or take 2%.2 Supply commitments have climbed to $279 billion, largely for memory tied to Vera Rubin production.6

Because the share price has not kept up, the valuation has come down sharply. Coverage uses different yardsticks, but they all point the same way. One analysis puts Nvidia at about 15 times expected fiscal 2028 earnings.2 Bloomberg-sourced reporting puts it at under 18 times forward 12-month earnings. That is close to a ten-year low and below the S&P 500's 19 times.8 On a trailing basis, one estimate has the multiple falling from above 50 in 2024 to about 28.3 Jensen Huang has started calling Nvidia "the world's first and only growth value stock."810

The market has also had to take in a lot. The stock was down 11% for the year on March 30, when investors questioned the return on AI infrastructure spending.8 A two-month slide before a late-July bottom wiped out more than $1 trillion in market value.1 The recovery picked up when Nvidia added $150 billion to its buyback program, the largest expansion in its history.19 Excitement about AI agents, including Meta's Muse, also helped.110

In one respect, Nvidia is still behind. The Philadelphia Semiconductor Index is up 87% this year, with Micron, Marvell and Intel each up more than 200%.8 By another measure, Nvidia's 28% gain trails a 96% rise in the iShares Semiconductor ETF.6 Bulls read this as a leader that still has to catch up with its sector.6 A less generous view is that the sector's biggest stock is also the hardest to move.8

Rubin and the falling cost of a token

The valuation question matters most for the products Nvidia ships, especially Vera Rubin, which is now ramping toward hyperscalers and neoclouds in the second half of 2026.14 Nvidia's main claim is about economics rather than speed: up to 10 times lower inference cost per token than Blackwell, and one-quarter as many GPUs to train mixture-of-experts models.1320 Each Rubin GPU is rated at 50 petaflops of NVFP4 inference, with 288GB of HBM4 and 22 TB/s of memory bandwidth, nearly three times Blackwell's bandwidth.1120

Nvidia has gone further. A company blog post updated in mid-September cites SemiAnalysis AgentX data from agentic coding runs. That data shows Vera Rubin NVL72 delivering 30 times more throughput per megawatt and 45 times lower token costs than GB300 NVL72.12 Neutral analysts are more cautious. One infrastructure provider says the 10x figure applies to mixture-of-experts and long-sequence work at scale. It adds that some analysts put gains for dense, short-context inference closer to two or three times.20 Pricing is also unclear. One reference guide says Nvidia has published no official Rubin price for the chip, board or system, and that dollar figures in circulation come from analyst models.19 Those models put an NVL72 rack at roughly $3.5 million to $4 million.11

My read is that the 10x claim is real for the workloads Nvidia designed Rubin for: agents, long context and mixture-of-experts models. It is also clearly a marketing ceiling. The 45x figure is best treated as a best-case result on a favorable benchmark. For buyers, the useful number is cost per token on their own workload, as one vendor argues.20

That matters because inference now dominates spending. One GPU cloud operator estimates that inference accounts for 70% to 80% of GPU cloud spend for teams that have shipped products.18 As a reference point, it cites on-demand B200 rates of about $1.03 per million tokens.16 Agentic workloads add pressure: OpenRouter data cited by Nvidia shows they use about 15 times more tokens than a simple chat request.12 If cheaper tokens lead to more agent use, Nvidia's efficiency gains increase demand instead of cannibalizing it. That is the bet behind the $1 trillion in combined Blackwell and Vera Rubin orders Huang has cited through 2027.11

The other side of rapid improvement

There is a tension here that most stock-focused coverage does not address. The faster Nvidia lowers the cost per token with each generation, the faster older chips lose economic value. This has become a financing problem. Nvidia's plan with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR aims to mobilize more than $500 billion for AI infrastructure.25 It relies partly on chip-backed loans. Lenders are pushing back. Banking sources told Reuters that some financiers want stronger guarantees than Nvidia first proposed, because they doubt the chips can serve as long-term collateral.29 Nvidia argues that top GPUs can earn revenue for up to a decade, and one valuation firm estimates a 9-to-10-year useful life for GB300 NVL72 systems.29 Lenders point to a lack of historical data and fast generational change as reasons for caution.30

Demand for financed compute is still huge. SpaceX is reportedly trying to raise $40 billion, about $10 billion in bank loans and $30 billion in investment-grade debt, to buy Nvidia chips. Apollo is expected to lead, and the talks are described as preliminary.2528 Amazon is reportedly looking at moving about $8 billion of Nvidia hardware into an investor-funded vehicle that it would lease back.26 Morgan Stanley estimates AI infrastructure will need $1.5 trillion in outside financing by 2028.25

Nvidia is also spreading out the places where its silicon is used. Microsoft's Surface Laptop Ultra, built on Nvidia's RTX Spark chip, starts at $2,599 and ships October 16. It is pitched as a way to run large models locally without cloud bills.24 Separately, an Nvidia-backed startup, Upscale AI, launched networking that links chips from rival vendors over open standards.27

The verdict

The historical record supports a narrow conclusion. Crossing $6 trillion will not decide Nvidia's next year. The race between earnings growth and multiple compression will.4 Some bullish forecasters see the stock nearly doubling in 2027 if earnings estimates hold and the multiple returns to market average.3 That scenario requires AI spending to keep growing. More than 90% of revenue comes from data centers, so a pullback in capital spending would hit earnings quickly.2 David Bahnsen's warning is worth keeping in mind: past capex booms burned a lot of capital before paying off, and he does not think AI has reached that stage yet.6

On balance, I think the stock's 2026 underperformance relative to its own fundamentals is more likely to narrow than to widen, because earnings are compounding faster than the multiple is shrinking. The bigger risk is not demand for Rubin. It is whether financing markets keep valuing Nvidia hardware the way Nvidia does, while each new generation makes the previous one cheaper to replace.