Anthropic has begun its IPO roadshow and disclosed a number worth watching: when it first connected Claude to AMD GPUs this year, the model completed the port in just 3 days, with nearly 100% of the kernels automatically written or ported by the model itself. Anthropic’s compute competition is shifting from simply purchasing chips toward using models to optimize how efficiently those chips operate.
What this is
Anthropic says Claude’s training and inference now span four major chip platforms: AWS Trainium, Google TPU, NVIDIA GPUs and AMD GPUs. Each chip requires its own software and inference toolchain. In the past, introducing a new platform typically required dozens of engineers working for months; this time, with Mythos participating in the port, the cycle was cut to three days.
The company also introduced the concept of “effective total compute”: physical compute multiplied by the efficiency gains discovered by the model. If an optimization reduces required compute to one-quarter, the same hardware can train a model of equivalent quality, or train a stronger model that previously required four times as much compute.
Industry view
The value of this approach extends beyond reducing reliance on a single supplier. Multi-chip capabilities can ease constraints on NVIDIA’s supply, pricing and delivery, while also allowing AWS, Google and AMD to compete for model companies’ compute orders. If models continue to improve utilization across different hardware, chip competition will expand beyond parameter scale and peak performance to include compilation, scheduling and inference costs.
But the risks are equally clear. The three-day port and automatically generated kernels Anthropic cites are based primarily on internal practices and have not been independently verified. It remains uncertain whether autonomous optimization systems can operate reliably over the long term. Roadshow materials inherently serve an IPO narrative; the three-day result and the claim of “recursive self-improvement” are better viewed as signals of competitiveness than publicly proven financial advantages.
Impact on regular people
For enterprise IT: When procuring AI services, companies can compare cloud platforms and chip architectures simultaneously, reducing the cost of becoming locked into a single GPU supplier. In the short term, however, the complexity of model migration and toolchains will remain within suppliers’ operations.
For individual professionals: Models participating in hardware optimization means AI’s value is extending beyond answering questions into R&D and infrastructure work. Knowledge workers’ competitive advantage will increasingly center on defining tasks, validating results and taking responsibility for decisions.
For the consumer market: Higher compute utilization may push down model API prices, making free or low-cost AI services more widespread. At the same time, the IPO narrative will accelerate efforts across the industry to translate technical progress into higher subscription fees or enterprise contract values.