NVDA Gained 10%. Here Is Why It Holds.

The Week in One Number

Nvidia closed Friday at $223.96. That is a 10.47% five-day gain, the stock’s strongest weekly performance since April, and its highest close since early June. The week opened with fresh hyperscaler capital expenditure disclosures, accelerated on Tuesday when SpaceX used the word exclusively on a live earnings call, and absorbed a Thursday TrendForce report on Rubin Ultra memory reductions without flinching. Two steps forward, one wobble that did not stick.

The record, however, is still out there. Nvidia’s all-time closing high of $235.47, set May 14, 2026, sits roughly 5% above Friday’s close. That gap is not indifference. It is the market doing specific arithmetic: the company posted $215.9 billion in annual revenue last fiscal year, and investors are now paying for whether the next annual figure clears $330 billion or stalls. August 26 is when the quarterly data that drives that answer arrives. Q2 FY2027 guidance was set at approximately $91 billion. The question is what lands above that number and what management says about Q3.

Three Events. One Direction.

Nothing about this week was random. Three distinct catalysts converged, and each one carries weight beyond a single trading session.

SpaceX said exclusively. On August 4, alongside SpaceX’s first-ever public earnings report, the company announced a joint development agreement with Nvidia to build the Starmind AI1 satellite compute payload: an orbital system carrying Vera Rubin NVL72-class hardware, including 72 Rubin GPUs and Vera CPUs, designed to run AI workloads from low Earth orbit. SpaceX envisions a constellation of up to one million such satellites. That alone would have moved the stock. Then Elon Musk went further. On the earnings call, Musk stated that SpaceX will build all of its AI infrastructure exclusively on Nvidia platforms, citing Vera Rubin as the best available option. He outlined a target of 2 gigawatts of compute capacity by end of 2025, scaling to 10 gigawatts by end of 2027. The word exclusively is doing significant work. It does not just signal a large order. It closes the door on AMD for an entire customer’s budget on an open-ended timeline.

The hyperscaler capex cycle did not slow. It accelerated. Aggregate 2026 capital expenditure guidance from the four largest U.S. cloud and AI infrastructure providers now totals approximately $725 billion: Amazon at roughly $200 billion, Microsoft near $190 billion, Alphabet in the $175 billion to $185 billion range, and Meta at $115 billion to $135 billion. That combined figure is up approximately 77% from the 2025 total of roughly $410 billion. The bulk is being directed at AI compute, data center construction, and GPU cluster buildout. Every dollar of that budget is, in large part, a dollar that converts into Nvidia hardware revenue. The cycle is not pausing.

The prior week’s selloff was a macro misread. Semiconductors were hit hard the week before on concerns about AI buildout durability and stretched valuations. The hyperscaler capex confirmations made that fear look exactly what it was: macro noise applied to a company-specific situation. When the data that follows a macro-driven selloff are positive and company-specific, the correction tends to be brief. This one was.

What Nvidia Actually Is in 2026

The word chipmaker still gets applied to Nvidia in casual coverage. It no longer fits. Nvidia is an AI infrastructure platform with hardware at the center, a software ecosystem as the moat, and a sovereign customer base that most semiconductor companies cannot access.

The current product generation is Vera Rubin. The NVL72, a rack-scale configuration of 72 Rubin GPUs paired with Vera CPUs, is already the reference architecture for large-scale AI deployments globally. Its competitive advantage is not only raw performance. The cable-less compute tray design cuts assembly time and total cost of ownership versus prior generations, which makes competing architectures harder to justify on a procurement spreadsheet even when their specifications look comparable on paper. SpaceX’s selection of this exact system for orbital deployment is not a marketing endorsement. It is a procurement decision by a technically sophisticated buyer with access to every available alternative.

CUDA is the lock that almost never makes headlines. AMD, Intel, and custom silicon from the hyperscalers themselves have all made progress on raw GPU performance. None of them has solved the accumulated software investment: the libraries, frameworks, tools, and institutional workflows that enterprises, research institutions, and cloud providers have built on CUDA over the past decade. Switching is not a procurement decision. It is a re-architecture decision measured in engineering years. That is what sustains Nvidia’s pricing power through each product cycle, not the chip specifications alone.

Sovereign AI deserves a separate line. Revenue from government AI infrastructure programs exceeded $30 billion in fiscal 2026, more than tripling year-over-year. Governments purchasing compute at that scale are not optimizing for cost. They are purchasing strategic independence from foreign technology supply chains. That category does not contract when the next chip generation arrives. It expands.

The Numbers Before August 26

• FY2026 full year: Revenue of $215.9 billion, up 65% year-over-year. Q4 revenue reached $68.1 billion, up 73% versus the prior year and 20% sequentially.
• Q1 FY2027: Revenue of $81.6 billion, up 85% year-over-year. Data center revenue surged 92% year-over-year to $75.2 billion, representing approximately 92% of total quarterly revenue.
• Q2 FY2027 guidance: $91 billion, plus or minus 2%, with gross margins guided to approximately 75%. Management noted the guidance assumes zero data center compute revenue from China, as H200 export approval remains unresolved.
• Analyst consensus: Average price target near $314. Full-year FY2027 consensus revenue estimates cluster in the high $300 billions range.
• Profitability: Return on equity of 114.29%. Debt-to-equity ratio near 0.07.
• Shareholder returns: Nvidia returned $41.1 billion through repurchases and dividends in FY2026. As of Q4, $58.5 billion remained under its buyback authorization.

Is the Valuation Still Defensible?

The question reduces to one thing: whether FY2027 consensus estimates are real. Nvidia’s earnings have outrun the stock price repeatedly, compressing the forward multiple faster than the share price rises. AMD still trades at a materially higher forward earnings multiple than Nvidia despite a fraction of Nvidia’s revenue base. That comparison matters because it frames what investors are actually paying for: Nvidia’s multiple is not cheap in absolute terms, but it is not stretched relative to a company growing data center revenue at 92% year-over-year with a 114% return on equity.

The Street is broadly constructive heading into August 26. What the consensus does not fully price is the China variable. Zero H200 revenue is already in the guidance. Any policy shift in either direction, approval or further restriction, is an event that re-rates the stock intraday. It is unmodeled in either direction.

Bull / Base / Bear

Bull case: August 26 delivers revenue above $93 billion, management guides Q3 above $97 billion, and Rubin Ultra HBM resolution arrives before shipments are affected. The SpaceX orbital compute program becomes a proof point that Nvidia hardware is moving into a category with no terrestrial competitor. The stock closes the gap to the May 14 all-time high quickly, then pushes into the $250 to $300 range where analyst price targets are concentrated.

Base case: Q2 revenue lands at $91 billion to $94 billion, gross margins hold near 75%, and Q3 guidance comes in at $94 billion to $97 billion. Rubin Ultra ships in a reduced-memory configuration that customers accept because the alternative, waiting for a chip that does not yet exist at scale, is worse. AI compute capacity remains constrained across 2026. The stock grinds toward the prior all-time high over two to three months.

Bear case: The HBM situation is structurally worse than the base case assumes. TrendForce has documented that since Q3 2026, Nvidia has been evaluating multiple downgraded HBM configurations for Rubin Ultra: moving away from the original 12-Hi HBM4E baseline toward 8-Hi HBM4E, 12-Hi HBM4, and 8-Hi HBM4 options, with no final specification confirmed. The mainstream SKU under evaluation would carry 192GB at 8-Hi, versus the existing Rubin chip’s 288GB at 12-Hi. This is not isolated. It follows Nvidia’s earlier decision to halve the SOCAMM capacity on its Vera Rubin Superchip modules due to LPDDR5X supply constraints. If the memory reduction is severe enough to measurably reduce training throughput per chip, hyperscalers could defer orders, gross margins could compress below the 74% threshold, and the stock retraces to $200 to $210.

Action Plan

Eighteen days to August 26. The decision framework is position-dependent.

No position: SpaceX’s exclusive commitment and the $725 billion hyperscaler capex cycle are not one-week signals. They are multi-year demand anchors. Entering now carries two-sided event risk on August 26. A scaled approach, half at current levels and half on any pullback toward $210, manages that risk without abandoning the thesis entirely.

Long with a gain: Trimming requires a specific bear thesis, not general market caution. The Rubin Ultra memory situation is the only structural argument with real weight. Absent a conviction view on that specific risk, hold is the default.

Long from a higher entry: The gap to the May 14 record close of $235.47 is under 5%. The business has grown materially since that peak. Patience is the correct position.

August 26 Scorecard

• Revenue vs. guidance: Q2 guided to $91B. A beat above $93B is the number that moves the stock. Watch this first.
• Gross margin: Guided at approximately 75%. Any reading below 74% is a yellow flag on HBM input costs.
• Data center concentration: Data center was $75.2B of $81.6B total in Q1, roughly 92% of revenue. That concentration is simultaneously the thesis and the single-point risk.
• Q3 guidance: Street expectations cluster in the mid-$90 billions. Below $93B disappoints. Above $97B re-rates the stock.
• Rubin Ultra HBM specification: No final configuration confirmed. Any clarity on August 26 changes the forward cost model immediately.
• Sovereign AI commentary: Over $30B in FY2026, up more than 3x. Listen for any FY2027 trajectory language on this line.
• China H200 policy: Zero revenue assumed in guidance. Any update in either direction is an unmodeled event.
• SpaceX Starmind schedule: Prototype launches targeted for early 2027. Any acceleration or delay changes the near-term orbital compute revenue math.
• HBM cost trajectory: DRAM supply is expected to remain tight through 2027. Watch the cost-of-revenue line against the 75% gross margin guide.
• AMD rack-scale progress: The competitive variable worth tracking into Q4. It has not changed the near-term picture. It matters for the twelve-month horizon.

Research Conclusion

The 10% gain was earned on identifiable, specific catalysts. SpaceX committed exclusively to Vera Rubin, on the record, with a 10-gigawatt compute target and an orbital satellite program attached. The four largest hyperscalers confirmed they are collectively spending $725 billion on AI infrastructure in 2026, up 77% from last year. The stock, after all of it, still trades 5% below its all-time high, with a forward multiple that compresses faster than the revenue grows.

The thesis into August 26 is precise. Revenue at or above $91 billion and a Q3 guide above $93 billion closes the gap to a new high quickly. A gross margin miss or a weak guide sends the stock back to $200 to $210. Both outcomes are live.

The Rubin Ultra memory situation is the only bear argument that carries structural weight. DRAM scarcity is now influencing chip specifications on the world’s most anticipated AI hardware, not just its price. That is a different category of risk than a supply delay. TrendForce’s documentation of the same pattern across SOCAMM, RDIMM, and HBM makes it harder to treat as isolated. Keep it on the list and watch the cost-of-revenue line on August 26.

Everything else favors owning NVDA into the print with a defined reassessment point. Hyperscaler spending is accelerating, not decelerating. The software moat is intact. SpaceX went exclusive. Eighteen days.