Ramez Naam proposed a shortcut around one of NVIDIA’s strongest defenses: ask an AI to take code written for its hardware and make it run fast on a rival’s chip. Another panelist immediately connected that possibility to what he described as $5 trillion in U.S. market value hanging in the balance.
Speaking on Moonshots with Peter Diamandis, Naam predicted that NVIDIA’s software moat—an advantage competitors struggle to cross—would break “this year and next year,” putting his forecast in 2026–2027. He also disclosed a financial interest: Lemurion, a company in his Planetary VC portfolio, is working on the problem.
Most of the conversation concerned electricity. This stretch concerned another part of the data-center bill: why buyers who want alternatives to NVIDIA have found switching difficult, and what might finally make it worthwhile.
The moat is more than the silicon
Naam described NVIDIA as a network-effect business that people rarely recognize as one. In his assessment, AMD’s chips were already competitive. What kept customers with NVIDIA was CUDA, its programming platform for getting GPUs—the processors widely used for AI—to do useful work.
The more developers build and optimize software around that platform, the more attractive it becomes to keep buying hardware that runs the software well. A competing chip has to overcome not just a hardware comparison but the work already invested in making NVIDIA systems fast.
Part of that work lives in small pieces of code called kernels. A kernel carries out a particular computation, such as multiplying arrays of numbers. An optimized kernel arranges that work to suit a chip’s processors and memory. Getting code to run on different hardware is one task; getting it to run efficiently enough to justify switching is another.
That is the rewriting cost Naam expects AI to reduce. Rather than have specialists manually adapt performance-critical code for each vendor, developers could ask an AI system to generate or rework it for the target chip. Compilation translates software into instructions the hardware can execute; the proposed breakthrough is automating more of the optimization needed to make those instructions fast.
Naam said Lemurion was working in this direction. Another panelist offered his own recent experience: he had had good luck using an AI model to produce custom kernels over the preceding couple of weeks, which made Naam’s prediction seem plausible to him. The exchange supplied a forecast and an encouraging personal example, rather than a measured comparison showing that entire production workloads could already move between vendors without a performance penalty.
The counterargument: the links between chips
The panel’s objection was that NVIDIA sells more than processors and a programming platform. It also supplies interconnect: high-speed links that let chips exchange information while working on a shared task. Naam called NVIDIA’s chip-to-chip connections strong.
Those links are especially valuable during training, when a model learns from data. In large distributed training runs, many chips divide up the work and repeatedly exchange results. Slow communication can leave expensive processors waiting rather than calculating.
Dave Blundin’s counterargument was that more of the work was moving to inference: using an already-trained model to answer a question, generate code or produce another output. He put the shift at 95% of the load and argued that inference did not really need the same interconnect.
The response from the table was narrower. A large model can still run across multiple GPUs, which need to communicate quickly. The example offered was roughly 10 to 20 GPUs working together within a rack, rather than a smaller, speed-focused model that might have different requirements.
A third panelist summarized the distinction as local versus global coordination. In the inference setup they were describing, the nearby chips serving a model must work closely together, but an entire supercluster need not stay coordinated in the way a large training run does.
The 95% figure was a panel estimate, not an established industry-wide workload split. Likewise, the rack and GPU count illustrated one kind of deployment, not a requirement for all inference. The competitive argument was about the scale of communication: a shift toward serving models could change where NVIDIA’s networking advantage matters most without making that advantage disappear.
Why buyers would bother switching
Blundin had introduced the economics by challenging the idea that GPUs are inherently so expensive that they must never sit idle. He estimated that electricity accounted for about 5% to 10% of a data center’s cost, then argued that hardware and AI-service prices contained substantial margins at several stages of the supply chain.
His rough calculation stacked what he called an 80% NVIDIA markup with twofold markups at the chip-manufacturing and model-provider levels. He described the eventual price as twenty times the cost of turning sand into a chip. These were conversational estimates spanning different businesses, not an audited breakdown of a GPU purchase price. Naam agreed with the broader point: “there’s a lot of margin going in there.”
That argument explains the incentive to make code portable. A cheaper competing processor is not necessarily a bargain if moving to it requires extensive engineering or makes the application slower. If AI reduces that conversion work while preserving performance, buyers gain more credible alternatives—and more leverage when choosing a supplier.
The potential saving is therefore not simply the difference between two chip prices. It depends on how much useful work each system can deliver after the software has been adapted, including the communication the workload still needs.
Two different kinds of compatibility
Earlier coverage of NVIDIA’s financing plans examined the other side of CUDA compatibility. In the preceding Moonshots discussion, Emad Mostaque described how software compatibility could help older NVIDIA chips keep earning on suitable workloads. Keeping code useful across NVIDIA generations supports the installed base; making it portable across vendors could give buyers a way out of it.
The existing idea page, The inner loop: why AI advantage may sit outside the model, develops a complementary argument from that earlier conversation: organizations may distinguish themselves through their data, workflow knowledge and ability to learn from deployment, even when capable models become widely available. Naam’s forecast concerns a different layer—the cost of changing hardware—not evidence that the organizational-learning thesis has been established.
For chip buyers, the practical question is concrete: can the same workload run correctly and fast enough on another vendor’s processors that the saving survives the move? AI-written kernels could lower the software barrier. The chips would still have to do the work, and the links between them would still have to keep up.