This One Metric Is Why I’m Going to Keep Buying Nvidia Hand-Over-Fist

One stubborn number buried in Nvidia's latest earnings report signals something that almost never survives at hardware scale, and it is the reason this investor keeps adding shares despite a valuation that looks stretched on the surface.

Published September 3, 2026, 11:30am ET · 3 min read

A close-up shot of a hand holding a smartphone. The phone screen displays a stock trading application for NVIDIA, featuring the bright green NVIDIA logo, a prominent green 'BUY' button, a red 'SELL' button, and a candlestick stock chart with red and yellow bars. The background is a blurred, out-of-focus blue display showing luminous numbers, hinting at market data.
A smartphone displays NVIDIA's stock trading interface with 'BUY' and 'SELL' options, reflecting strong investor interest, particularly given the company's impressive gross margins. © Shutterstock / rafapress

I keep hitting the buy button on NVIDIA (NASDAQ:NVDA | NVDA Price Prediction) because one number refuses to break, and that number tells me everything I need to know about pricing power in this cycle.

That number is gross margin. At the scale NVIDIA now operates, holding a 75% non-GAAP gross margin should be arithmetically difficult for a hardware company. The Q2 FY27 report landed exactly there: revenue of $96.22B, gross profit of $72.14B, non-GAAP gross margin at 75.0%. Management said the level was “largely unchanged from last quarter due to a similar product mix.” When a company sells $89 billion of data center gear in ninety days and still keeps three quarters of every dollar as gross profit, that is monopoly economics at hardware scale.

Pricing Power That Refuses to Compress

Q2 FY27 revenue grew 105.85% YoY. Data center revenue reached $89.02B, up 117% YoY. Every one of those dollars was harder to earn than the one before it, because supply chain constraints hit foundry, memory, power, and yield simultaneously. Margins still held at 75%. Management guided Q3 gross margin to 74%, plus or minus 50 basis points, and fiscal 2028 to 72% to 73% once executed price increases work through memory costs. Even that compressed figure keeps NVIDIA well above any data center silicon peer. Vera Rubin, now in full production, expands the revenue opportunity to $40 billion per gigawatt, up from Blackwell’s $25 billion and Hopper’s $18 billion. Customers pay more because the platform delivers more.

Returns That Compound Faster Than the Multiple

Return on invested capital is 92.2%. Return on equity is 101.5%. Debt-to-equity sits at 0.073, with interest coverage of 503.4x. Free cash flow in fiscal 2026 was $96.58B, up 58.7%. In Q2 alone NVIDIA returned $26 billion to shareholders, $20 billion in buybacks and $6 billion in dividends, with roughly $99B still authorized. Trailing P/E of 45 looks full, yet consensus already models fiscal 2028 EPS of $13.1277 against $4.77 in FY26. When earnings compound this way and capital returns scale with cash flow, the multiple resolves itself.

NVIDIA’s Edge Over Broadcom, AMD, and Amazon’s Trainium

Broadcom (NASDAQ:AVGO) sells custom silicon into specific hyperscaler sockets. Advanced Micro Devices (NASDAQ:AMD) builds a credible accelerator. Amazon (NASDAQ:AMZN) has disclosed Trainium as a multibillion-dollar business. NVIDIA’s data center segment printed $89.02B in a single quarter. A multibillion-dollar Trainium line is a rounding item inside one NVIDIA reporting period. Intel (NASDAQ:INTC) competes on chips too, and now collaborates with NVIDIA on NVLink custom products. NVIDIA sells the CUDA software estate, NVLink fabric, Spectrum-X networking, and the Vera CPU on top of the accelerator. Networking grew 18% sequentially, with SpectrumX Ethernet up 2.6X year over year. That is the moat you cannot backfill by taping out one accelerator.

Risks on My Radar

China data center compute revenue is fully excluded from forward guidance. Supply obligations sit at $279B, guarantee obligations reach $108.5B, and days sales outstanding stretched from 45 to 60 days as investment-grade customers took extended payment terms on multi-quarter shipments. If the AI capex cycle rolls over, those commitments become the headline. What keeps my capital active is that cloud-industry backlog exceeds $2 trillion, and top-five hyperscaler capex is guided at nearly $800 billion in 2026 and $1.3 trillion in 2027. All of that spend has to be powered, cooled, and networked by somebody, and we rounded up seven suppliers behind the buildout in a free report on the AI boom beyond the chipmakers. Demand is running well above what NVIDIA can supply at least through the end of fiscal year 28.

Why the Buy Button Stays Active

Jensen Huang said it plainly on the call: “Now, compute is revenue.” Until the gross margin line breaks, my capital keeps landing on the same ticker.

Contact [email protected] for any questions or corrections.

Alex Sirois

Alex Sirois is a financial writer with experience spanning both retail and institutional investing. He has written for InvestorPlace and held roles at BNY Mellon and Bernstein, giving him a perspective that bridges Main Street portfolios and Wall Street analysis.

Alex holds an MBA from George Washington University and has built his career across multiple industries, including e-commerce, education, and translation — a breadth of experience that informs how he breaks down complex financial topics for everyday investors. His writing is conversational, actionable, and grounded in long-term, buy-and-hold investing principles.

At 247 Wall St., Alex focuses on delivering analysis that is both accessible and useful, with a clear emphasis on helping readers make more informed decisions with their money.

All articles →