Nvidia AI Investment Case Holds Up, But Margin Pressure Is Building

Nvidia’s AI investment case remains intact heading into the second half of 2026, but the numbers that underpin it are becoming more complicated. Revenue is still surging, the competitive moat in GPU software is real, and hyperscaler spending shows no sign of pulling back. The margin line, however, is beginning to move in the wrong direction, and that deserves more attention than it typically gets in the bull-case narrative.

The Nvidia AI Investment Case in Numbers

NVIDIA’s full-year Fiscal 2026 revenue rose 68% to a record $193.7 billion, with fourth-quarter revenue of $62.3 billion, up 75% year-on-year. The quarter before that, Q3 FY2026 delivered $57.0 billion, up 62% from a year earlier. Those are not the numbers of a company losing its grip on a market.

The most recent data extends the run further. According to REX Shares’ analysis of NVIDIA’s Q2 FY27 disclosures, Data Center revenue reached $89.0 billion in that quarter, representing roughly 92% of total revenue, with gross margin at 75.0%. The company guided Q3 FY27 revenue to $108.0 billion. For context, full-year Fiscal 2025 revenue was $130.5 billion. NVIDIA is now guiding a single quarter above that entire year.

The demand side is not the concern. The same REX Shares analysis notes that Microsoft, Alphabet, Amazon and Meta spent a combined $166.0 billion on capital expenditure in the most recent reported quarter, up 87% year-on-year and 27% sequentially. Over the ten quarters since the start of 2024, combined hyperscaler capex has risen 272%. That spending flows, in large part, through NVIDIA’s order book.

Where the thesis gets tested is on margins. In Q4 FY2025, NVIDIA’s GAAP gross margin was 73.0%, down 3.0 percentage points year-on-year. By Q2 FY27 that had recovered to 75.0%, but the directional pressure from rising memory and component costs, flagged repeatedly in company filings, has not gone away. Microsoft, Alphabet, Amazon and Meta are all developing proprietary AI chips. AMD is gaining credibility as a GPU alternative. As in-house silicon matures, the willingness of hyperscalers to pay premium prices for NVIDIA hardware will face a genuine test.

CUDA and the Networking Edge

The single most durable element of the Nvidia AI investment case is not the hardware. It is CUDA (Compute Unified Device Architecture), the software layer that connects AI applications to NVIDIA GPUs. According to Business Insider, CUDA has millions of lines of code, is optimised natively for more than 250 frameworks including PyTorch and TensorFlow, and at one point gave NVIDIA around 90% market share in AI computing. Amazon’s own internal documents, obtained by Business Insider, noted that Amazon’s Neuron software “prevents migration from NVIDIA CUDA,” which they identified as a primary factor holding back AWS customers from switching to Amazon’s own chips.

That is an extraordinary admission from a direct competitor. It also illustrates why Mizuho analysts wrote: ‘We continue to see NVDA leading the AI race with its fortress CUDA moat and strong hardware/networking performance for AI training and inference.’ Analyst Patrick Moorhead offered a more qualified view, arguing that CUDA remains a moat but a shrinking one, and that NVIDIA’s balance sheet and the customer financing it enables may now be the larger structural advantage.

On the hardware side, NVIDIA’s Blackwell Ultra architecture claims up to 50x better performance and 35x lower cost for agentic AI workloads compared with the Hopper platform, based on SemiAnalysis InferenceX benchmarks cited in the Q4 FY2026 earnings release. Networking is becoming a second moat: third-party analysis from TradingKey estimates NVIDIA’s networking segment at roughly $8.2 billion, or about 15% of Data Center revenue, growing at 162% year-on-year.

Challenges to CUDA’s dominance are forming, slowly. In September 2024, Qualcomm, Google and Intel formed the UXL Foundation to build a chip-agnostic software rival. Progress has been gradual. A software ecosystem built over fifteen years does not unwind quickly.

On the supply side, secondary analysis citing NVIDIA’s own disclosures puts purchase commitments for foundries, memory and packaging at approximately $100 billion, with some estimates running considerably higher. That scale of forward commitment is both a sign of confidence and a source of execution risk if demand softens.

At last check, NVIDIA carried a market capitalisation of $5.769 trillion, a trailing PE of 30.20, EPS of $7.91, and a five-year return of 1,073.30%. The consensus analyst price target was $328.72, against a 52-week range of $164.27 to $240.10. The valuation reflects an expectation of continued dominance rather than mean-reversion.

NVIDIA has also authorised an additional $150 billion in share repurchases, signalling that the board views the current price as manageable relative to the cash the business generates. That is not a defensive move; it is a statement of conviction about the earnings trajectory.

The next real test for the Nvidia AI investment case arrives when hyperscaler custom silicon reaches production scale. If in-house chips begin displacing NVIDIA GPUs at the margin, it will show up in gross margin and Data Center revenue growth before it shows up in the headlines. The Q3 FY27 revenue print, against guidance of $108.0 billion, will be the first meaningful read on whether that displacement has begun.