Dylan Patel on the AI Chip Race – NVIDIA, Intel & the US Government vs. China

“How you buy GPUs is like buying cocaine. You call up a couple of people, you text a couple of people, you ask, yo, how much you got? What’s the price?” This colorful analogy from Dylan Patel, Chief Analyst at Semi Analysis, underscores the frantic, opaque, and high-stakes reality of procuring AI infrastructure today. It sets the tone for a wide-ranging discussion that dissects the seismic shifts rocking the semiconductor world, from NVIDIA’s shocking $5 billion investment in Intel to the relentless rise of Huawei as a formidable competitor. The conversation explores not just the tactical moves of chip giants but the broader geopolitical and economic forces shaping the future of AI compute.

At the heart of the discussion is the unexpected alliance between long-time rivals NVIDIA and Intel, described as a “poetic full circle” and a “Buffett effect” for the chip industry. Analysts see this as a lifeline for Intel and a strategic masterstroke for NVIDIA, potentially creating an “x86 laptop with NVIDIA graphics fully integrated” that could be the best product on the market. The move is seen as catastrophic for competitors like AMD and ARM, whose positions are suddenly undermined by this team-up of “arch nemesis” companies. The deal symbolizes a fundamental remixing of the competitive landscape.

The analysis then pivots to the East, with a deep dive into Huawei’s alarming capabilities. Despite severe U.S. sanctions, Huawei has not only survived but is innovating at a pace that worries NVIDIA CEO Jensen Huang himself. The discussion reveals how Huawei’s announcements—like custom high-bandwidth memory (HBM) and specialized chips for pre-fill and decode workloads—are part of a sophisticated “10,000 IQ” strategy to pressure the U.S. into allowing more exports, all while building a domestic supply chain that could eventually compete globally. The tension between China’s mandate for self-reliance and the practical needs of companies like ByteDance, which “just want the best chips” from NVIDIA, creates a volatile and uncertain market.

Finally, the conversation examines the evolving ecosystem of AI infrastructure, from the resurgence of Amazon Web Services driven by Anthropic to Oracle’s audacious, debt-fueled bet to become the backbone for OpenAI. It highlights the insane scaling of data centers, now measured in gigawatts, and the intricate hardware considerations for companies choosing between NVIDIA’s latest Blackwell GPUs. The underlying theme is a market in hyper-growth, where capital, engineering speed, and sheer will—epitomized by figures like Elon Musk building gigawatt-scale clusters in months—are the new currencies of power. The era of easy GPU procurement is over, replaced by a complex, high-stakes game that mirrors the transformative potential of AI itself.

Surprising Insights

• Huawei’s Announcements as Strategic Negotiation: China’s hyping of its domestic chip capabilities, like Huawei’s multi-year roadmap, is interpreted as a deliberate, high-level tactic to scare the U.S. government into relaxing export controls, under the logic that if China is already so advanced, there’s less risk in selling them better chips.
• NVIDIA’s Culture of “Cutting Features to Ship”: Beyond its software moat, a key to NVIDIA’s success is a relentless focus on time-to-market, enforced by internal figures who are “hated” by technologists for prioritizing shipping now over perfecting later. This cultural trait of betting the company on gut-feel volume orders is contrasted with the more cautious approaches of rivals.
• The “Galapagos” Risk for the West: An intriguing theory suggests that by restricting China’s access to cutting-edge tech, the U.S. might unintentionally force China down a different, more optimal technological path, while the West gets stuck in a “local minima” of its own design, optimized for current AI trends but potentially a dead end.
• Reliability as a Growing Bottleneck for Next-Gen Chips: With NVIDIA’s massive Blackwell GPUs (like the 72-GPU GB200), failure management becomes a major operational headache. Clouds are now offering SLAs for only 64 of the 72 GPUs, and companies need sophisticated “high-priority/low-priority workload” schemes to manage outages, making raw performance gains harder to realize in practice.
• Oracle’s Capital Structure as a Competitive Weapon: Oracle’s rise as a cloud contender is less about superior tech and more about its willingness to use its massive balance sheet and debt capacity to underwrite gigantic, long-term infrastructure deals (like the one with OpenAI) that more conservative hyperscalers like Microsoft are hesitant to take on.

Practical Takeaways

• Diversify Your Cloud Strategy: Don’t rely on a single provider. The performance, pricing, and availability landscape across AWS, Google Cloud, Microsoft Azure, Oracle, and specialized “neo-clouds” like CoreWeave is changing rapidly. Conduct regular RFPs (or “drug deal”-style outreach) to find the best combination of cost, capacity, and specialized hardware for your specific workload.
• Architect for Pre-fill/Decode Disaggregation: For inference workloads, design your systems to separate pre-fill (context loading) and decode (token generation) operations onto different hardware pools. This allows for independent scaling, better guarantees on time-to-first-token, and prepares you for future specialized chips like NVIDIA’s upcoming CPX, which are optimized for pre-fill.
• Factor in Reliability and Total Cost for New Hardware: When evaluating next-generation GPUs like NVIDIA’s Blackwell, look beyond peak FLOPs. Calculate the Total Cost of Ownership (TCO) that includes potential downtime, the complexity of managing large, coherent GPU domains, and the actual performance gain for your specific models (which can vary from 2x to 6x+).
• Treat China as a Separate, Advanced Market: For any global strategy, recognize that China’s AI chip ecosystem, led by Huawei, is advancing rapidly despite sanctions. Assume they will solve manufacturing bottlenecks over time. This means products, partnerships, and competitive assumptions must account for a potential parallel, powerful tech stack emerging from China.
• Build Relationships and Scout Continuously: GPU procurement is not a one-time purchase but an ongoing intelligence operation. Build direct relationships with cloud account teams and niche providers. Stay informed on delivery timelines, new data center openings, and emerging hardware to secure capacity and favorable terms in a perpetually tight market.