GPU Prices Up 53%, $860B Chip Market & Xi-Trump Summit Silence

Jamie Cole (0:00)

Huawei Hardware & Chips, Daily News. I'm Jamie Cole, thanks for joining me. Today, the 2nm crunch, ATSMC's 20% order surge signals next bottleneck. TSMC's most advanced manufacturing node just got a lot more crowded. Apple, Nvidia, AMD, Qualcomm, and MediaTek have all increased their 2nm capacity orders by 10 to 20%.

That's not a modest uptick. That's a coordinated scramble for the same scarce resource at the same time. Here's what that signal tells us. When five of the largest fabless chip designers move simultaneously to locking capacity, it means they're not confident supply will be there if they wait. The 2nm ramp is ahead of internal expectations at TSMC, which is the good news. The bad news is demand is growing faster than the ramp. TSMC is also expected to raise 2nm pricing by 3-6% in early 2027 Customers know that locking in volume now is partly technical necessity, partly price hedging. Both motivations point to the same conclusion. The next manufacturing bottleneck is already forming. The clearest real-time confirmation of that supply squeeze came from Nebius. The sovereign AI cloud provider raised on-demand hourly rates for Nvidia H100, H200, B200 and B300 GPUs by 17% to 21%, effective October 1st. An H100 that cost $2.95 per hour in May now runs $4.50. That's a 53% increase in under six months on hardware that's already a generation old. The important distinction here is that this isn't a startup struggling to manage costs. Nebius is spending $20 to $25 billion on infrastructure and can reportedly sell its entire 2027 capacity today.

Raising prices while demand still exceeds supply at that scale tells you something about the structural gap between what the market needs and what currently exists.

Against that backdrop, consider where the market is heading. The data center AI chip market stood at $124 billion in 2024

By 2030, that figure is projected to reach $860 billion, growing at 38% annually.

The shift from training to inference is a meaningful part of that story. Inference hardware tends to be lower power and cheaper per unit, but the sheer volume of deployment is expected to more than compensate.

Here's the thing. Nvidia sits at 78.2% market share today. Google at 4.7, AMD at 4.1, Intel at 3.7, Huawei at 2.6.

The inference transition is the most credible structural threat to that dominance, particularly as Korean and Indian chipmakers target lower power and PU designs.

The signal to watch is where the inference fragmentation becomes commercially meaningful before Nvidia ports its own inference optimizations deeply enough to foreclose the opening. The geopolitical picture shifted too. The Xi-Trump summit produced no movement on semiconductor export controls. Nvidia and AMD CEOs were seated at the head table. The issue didn't make it into official talks. That absence is itself a data point. It could reflect backroom deal making, political deferral, or tacit acceptance that the current control regime is stable enough that neither side wants to reopen it formally right now. What's certain is that Nvidia's path back into the Chinese market remains blocked, and the lobbying leverage of the past week didn't move the needle. That creates an opening. FuriosaAI and Rebellions, the two most prominent Korean AI chip makers, are actively positioning for Middle East markets as a hedge against US export control risk. Middle Eastern governments are looking for non-Navidia options on cost, power efficiency, and supply chain control grounds. The Korean firms are three or more years behind Nvidia architecturally. That gap matters. But the window depends less on catching up to Nvidia, and more on whether the geopolitical pressure that creates demand for alternatives remains sustained. If Nvidia supply normalizes or controls ease, the window narrows quickly.

Two final markers worth tracking. TSMC and Samsung have both committed to ASML's high-NA UV lithography deployment by 2028 That deepens the equipment concentration that already defines advanced node manufacturing. Intel, still fighting to close the node gap, faces continued pressure as its largest competitors lock in next-generation tooling, and the risk the whole industry is quietly circling. The 2027 CAPEX sustainability question. Big Tech is spending over $1.1 trillion annually on AI infrastructure. Multiple analyses now flag that hyperscaler capital expenditure could exceed free cash flow generation in 2027 If that forces a spending pullback, the downstream impact on TSMC, ASML, memory makers and specialty material supplies would be significant. No one is treating that as a base case yet. But it's the stress test that the entire current investment cycle has to pass. The near-term watchpoints are clear. TSMC's 2nm yield trajectory, whether Nebius-style pricing spreads to other cloud providers, and any signal from Washington on export control direction after the summit's silence.

Those three data points will say more about where this market goes next than almost anything else. Thanks for listening. This podcast was built using AI technology, a YesWe production.