Nvidia’s Moat Is Closing: China Chips and the Memory Shift

Key takeaways

• Nvidia is still the world’s most valuable company at about 5.3 trillion dollars, data centre revenue up 92 per cent to 75.2 billion dollars.
• Nvidia’s market value rose about 17 per cent in 2026, from 4.53 to 5.31 trillion dollars. Micron’s rose about 215 per cent. The re-rating went to the bottleneck, not the brand.
• Nvidia’s share of China’s advanced AI accelerators went from roughly 95 per cent to zero, on Jensen Huang’s own account.
• The counter-fact: about five homegrown DUV machines ship in 2026, and CXMT’s prospectus carries no HBM project.

Nvidia’s position never rested on Nvidia being good. It rested on there being no credible alternative. That is a more fragile claim, and I have argued this since before it was consensus.

Alternatives are arriving from two directions. China is building its own accelerators, memory and lithography. Separately, the scarce input moved down the stack: not the GPU, but memory. This is about where margin sits in an AI buildout. It is not investment advice, and I am not a licensed adviser.

The business is fine. The margin moved.

April-quarter revenue was a record 81.6 billion dollars, up 85 per cent. Non-GAAP gross margin hit 75.0 per cent, up from 61.0 per cent. The honest caveat: that year-ago base carried a 4.5 billion dollar H20 charge, and excluding it the comparison is 75.0 against 71.3, so the headline jump overstates the underlying improvement. Operationally, nothing is wrong.

Two companies is not a sector claim. But this is what it looks like when the binding constraint moves. Markets pay for scarcity, not growth.

For that to reverse, memory supply must catch up with AI demand. It has not. AI data centres absorb an estimated 70 per cent of global memory output, and Gartner expects memory prices up about 130 per cent by end 2026.

Servers and GPUs take almost 60 per cent of a gigawatt-scale buildout, but memory sits inside that share and is rising, while facility, networking and power take the rest. Nvidia holds a shrinking slice of a growing pie, the same logic as the landlord economics of India’s AI buildout.

How the moat actually erodes

Each step is survivable alone. Together they compound, because demand denied to Nvidia funded its replacement. Same pattern as the TCS layoffs: strong numbers right until the model stops holding.

China now makes what it was missing

Source: China National Bureau of Statistics, H1 2026, via Xinhua

Read the 2,580 per cent for what it is. It is a base effect, and the base is not published: the statistics bureau gives a growth rate for integrated circuit manufacturing and no yuan figure at all. What it does publish is the parent industry, which earned 302 billion yuan in the first half of 2025 against 578 billion in the first half of 2026. Semiconductor discrete devices rose 31 per cent over the same period, so this is concentrated, not broad. CXMT raised 57.92 billion yuan, the largest listing in STAR Market history, closing 466 per cent up and displacing ICBC as China’s most valuable listed company.

The caveat: about five domestic DUV systems ship in 2026 and twenty in 2027, entering SMIC’s flow only from 2027, and Asia Times reports China still trails ASML by four generations. Real capability, slow clock. India’s own subsidy attempt, in the PLI reality check, shows how long these clocks run. Risk runs both ways: Micron fell 28.7 per cent in July 2026.

What I expect next

• Watch Nvidia’s gross margin, not revenue. Below the low seventies while revenue grows is the moat compressing.
• Memory stays tight through 2027. CXMT announces an HBM programme within eighteen months.
• The next re-rating goes to power and cooling, not silicon.

What to do with this

• Building on GPUs: stop assuming compute cost falls in a straight line. Memory inflation can hold cost per token flat.
• Procuring hardware: lock memory pricing further out than feels comfortable.
• Track the constraint, not the logo. None of this is a recommendation to buy or sell any security.

Read next: how AI+ booked 1,650 crore rupees despite terrible reviews, and why India’s funding drought is not a winter.

Sources

• Nvidia Q1 FY2027 results, 20 May 2026; Tom’s Hardware on Jensen Huang’s China remarks; Epoch AI on 1 GW data centre cost
• China National Bureau of Statistics H1 2025 and H1 2026 industrial profit releases, via Xinhua, SCMP and ecns.cn; TechNode, CNBC and TrendForce on CXMT; Asia Times on DUV; SCMP on DeepSeek V4-Pro; stockanalysis.com and companiesmarketcap.com for market capitalisation

Figures are as reported by the sources named above at the time of writing.