The Chip Shortage and the Software Land Grab Are the Same Bet

Nvidia's own researchers cannot get enough of Nvidia's own chips. Bryan Catanzaro, who runs the company's applied deep-learning research, said it plainly.

We’re all supply constrained. — Bryan Catanzaro

Jensen Huang personally decides which internal teams get GPUs ^[1]. And in the same stretch, the company that cannot make hardware fast enough spent $12.9 billion to buy Hugging Face, a repository of open-weight models — software, not silicon ^[2]. The company rationing chips is buying distribution. That is not a contradiction. It is a hedge, and it is the clearest single picture of where the industry's money is now going. The model layer is commoditizing fast. Dell reports token prices fell 80% year over year even as reasoning-token usage rose 320-fold ^[3]. DeepSeek cut its V4-Pro prices by 75%, making it 12 to 19 times cheaper per task than GPT-5.5 or Claude Opus 4.7 ^[4]. By August, OpenAI, Anthropic, xAI, and Meta were all slashing prices to counter Chinese rivals ^[5]. A model that was a moat two years ago is commoditizing fast. Meanwhile the chip moat is eroding. Custom ASICs are projected to reach roughly 28% of AI compute ^[6]. AI coding agents recreated CUDA-like software in ten hours — Infinity did it for D-Matrix ^[7]. Groq licensed its inference technology to Nvidia itself ^[8]. And Nvidia put $2 billion into Marvell, a custom-ASIC designer whose entire business is building the chips that would erode Nvidia's GPU dominance ^[6]. That is the hedge visible on a balance sheet. So the land grab is at the deployment layer. On April 24, every major lab launched an enterprise agent platform on the same day — OpenAI's workspace agents, Microsoft's Foundry, Google's Gemini Enterprise, Anthropic's 200-partner connector system ^[9]. Forward-deployed AI engineer postings are up 1,000% this year ^[10]. Anthropic is embedding Claude inside Excel, PowerPoint, and Slack rather than a separate window ^[11]. Snowflake is putting Claude directly into its governed corporate data, and its product chief Christian Kleinerman explained the appeal.

Customers want AI that works directly on their governed data, not in isolated systems. — Christian Kleinerman

Nvidia and Palantir are building an enterprise AI operating system ^[12]. None of this means the hardware race is over. It is still binding. Nvidia has committed $279 billion through fiscal 2029 to secure memory from SK Hynix and Micron ^[13], and Huang has said as much.

supply (not demand) is capping that growth. — Jensen Huang

Nebius reports demand still outstrips supply ^[14]. Musk names memory as the limiting factor ^[15]. The chips are still scarce. That is precisely why this is a hedge and not a pivot. The economic logic is the old one about complements: when the layer you dominate commoditizes, the value migrates to the layer next to it — the one between the model and the user. No single source says this. The pattern says it. The same companies running hardest to secure chips are the ones racing hardest to own what sits on top of them. They are running two races at once, not because they have chosen one over the other, but because winning only one would be losing.