Key Takeaways

  • Anthropic reported that its Mythos model ported itself to AMD hardware in three days, replacing six to twelve months of human engineering.
  • AMD CEO Lisa Su confirmed over the summer that AI speedups have slashed hardware bring-up times across the board.
  • Software moats disappear when automated code generation treats underlying silicon as completely fungible compute.
  • Nvidia faces the same trap Google faced: selling the compute that trains the very models built to disrupt its proprietary lock-in.
  • Dario Amodei and Anthropic are pushing hardware independence at a moment when Jensen Huang faces political friction.

The Three-Day Port

For nearly two decades, Nvidia built an iron wall around its chips. That wall was not merely the silicon; it was CUDA, the proprietary software layer developers had to write for if they wanted their models to run fast. Switching hardware meant hiring an army of specialized systems engineers to rewrite kernel code line by line. That process routinely took six to twelve months of grueling work.

That dynamic broke this week. John Coogan pointed to Anthropic's recent milestone: “Anthropic sees all compute is fungible and it only took Mythos three days to figure out how to run itself on AMD hardware compared to six to 12 months.” (play from 7:53)

Three days. A frontier model wrote the low-level translation layers itself. Coogan noted this matches what chipmakers are seeing directly: “When we talked to Lisa Su at AMD a couple months ago over the summer, this was the big story of the day was that the bring up time had been significantly reduced from these AI speedups.” (play from 8:02)

When code generation drops the cost of software translation to zero, developer lock-in evaporates. The hardware choice stops being an architectural commitment and turns into a spot-market auction.

The Google Trap

Jordi Hays argued that this outcome was obvious to anyone tracking coding models: “As you started to understand the implications of how quickly we would be able to generate software you had to be bearish on CUDA right the two things were completely at odds.” (play from 9:52)

The irony is structural. Nvidia earned record margins by selling the compute that trained these models. Now, those same models are dismantling the software moat that justified Nvidia's pricing power. Hays drew a direct comparison to search: “Nvidia is helping create its own disruption in the same way that you saw with Google creating the original LLM also creating the technology to ultimately disrupt itself.” (play from 10:05)

This software shift arrives alongside political friction. With Huang under pressure regarding an alleged Chinese spy scandal, Hays observed that Anthropic CEO Dario Amodei has room to maneuver: “Dario will be out there continuing to push Anthropic's worldview at a time when Jensen is weakened and his ability to fight for his view is diminished.” (play from 10:24)

If models can port themselves across architectures in seventy-two hours, compute becomes a pure commodity. The winner is not the vendor with the stickiest proprietary libraries; the winner is whoever delivers the cheapest floating-point operations per watt.

What to Do With This

Audit your model inference stack tomorrow morning. If your engineers are writing custom CUDA kernels directly to squeak out a 15 percent speed gain, stop them. Direct that team to test running automated translation across alternative hardware like AMD Instinct chips, because relying on proprietary software layers will price you out of the market within twelve months.