Anthropic tool makes switching from Nvidia AI chips faster and cheaper
A new software tool from Anthropic can port AI workloads to rival hardware in days rather than months, potentially weakening Nvidia's market dominance.
The short version
- Claude Mythos can port AI models to AMD's MI355X GPUs in about three days.
- The previous standard for moving AI workloads to different hardware was six to 12 months.
- Anthropic is projected to move to a multi-gigawatt scale deal with AMD by the first half of 2027.
- Nvidia's share of contracted future AI compute for frontier labs has fallen to around 39%.
Why it matters
For businesses, this means AI hardware is becoming a commodity where price and supply can dictate choices. It reduces the 'lock-in' effect that previously forced companies to stick with Nvidia's expensive ecosystem.
Breaking the software lock-in
Nvidia has long maintained a dominant position in the AI market because of its CUDA software ecosystem. This software is the collection of tools and libraries that developers use to build and run AI models on Nvidia's hardware. Because so much work was already built on CUDA, moving to a different chip meant rebuilding the entire underlying infrastructure.
Anthropic recently disclosed a tool called Claude Mythos that changes this dynamic. The software can rewrite and tune AI models to run on different hardware with almost no human intervention. This allows models to hop between platforms much more easily than was previously possible for engineers.
As reported by Crypto Briefing, this capability turns AI chips into interchangeable parts. When code can move between different hardware platforms cheaply, the specialized software advantage that Nvidia enjoyed begins to fade. Companies can now prioritize other factors, such as cost or availability, over specific brand loyalty.
Drastic reductions in migration time
The speed of this transition is the most significant change for the industry. Anthropic reported that Claude Mythos can port an AI model to AMD’s MI355X GPUs in approximately three days. Before this tool existed, the same task typically required six to 12 months of work from dozens of engineers.
This shift removes the primary barrier to entry for competitors like AMD and Alphabet. Previously, the high cost of time and talent made it risky for labs to experiment with non-Nvidia chips. Now, the technical friction of switching is significantly lower, making it easier for organizations to diversify their hardware sources.
The goal of this automation is to make the transition seamless for the end user. By handling the complex rewrites and tuning automatically, Anthropic is allowing its customers to be more flexible. This move comes as the company prepares for its initial public offering, signaling a shift toward more open hardware compatibility.
The shift toward lower costs
Cost is a major driver behind the move toward alternative hardware. AMD reportedly offers a per-gigawatt system cost that is roughly half that of some Nvidia platforms. For large-scale operations, these savings become substantial as they scale up to multi-gigawatt levels of power.
Anthropic is already moving toward these alternatives with a deal for AMD’s Instinct GPUs. The first phase of this agreement is set for the first half of 2027. This deal serves as a proof of concept that other buyers can use to justify moving away from Nvidia's premium pricing.
Market data suggests that the trend toward diversification is already happening. The portion of future AI compute capacity that frontier labs have contracted with Nvidia has dropped to roughly 39%. This is a notable decline from previous highs of over 80%, showing that labs are spreading their purchases to manage costs and supply chain risks.
Winners and losers in the ecosystem
Alphabet stands to benefit from this trend through its custom tensor processing units, or TPUs. These chips are designed specifically for AI workloads, and they become more attractive when customers can migrate models onto them without a costly rewrite. This expands the pool of potential workloads that can run on Google Cloud hardware.
TSMC may be the most consistent winner in this transition because it manufactures the actual silicon. Since TSMC produces chips for Nvidia, AMD, and Broadcom, it collects revenue regardless of which designer wins the contract. As long as demand for AI compute remains high, TSMC will benefit from the production volume.
For Nvidia, the immediate threat is not necessarily a total loss of sales. Instead, the risk is a long-term loss of pricing power as the market becomes more competitive. While Nvidia still holds a massive installed base and deep relationships with nearly every major buyer, the ability to switch hardware more easily puts pressure on their dominance.
Limitations and future outlook
While the ability to port models quickly is a major step, there are still hurdles to overcome. Anthropic's results currently cover only part of its workloads, not every single task. Additionally, porting a model is different from ensuring it runs at peak efficiency across every specific use case.
Nvidia still maintains a very large footprint in the industry with established relationships with nearly every major buyer. These deep ties and the maturity of the existing ecosystem mean that the transition will likely happen gradually rather than overnight. The industry is moving toward a model of multi-vendor procurement.
The next major milestone to watch is the first phase of the Anthropic and AMD deal in 2027. This will provide a clear indication of how well non-Nvidia hardware can handle frontier-scale workloads. As tools like Claude Mythos become more common, the AI industry will likely see a more diverse and price-sensitive hardware landscape.
What to watch
- The first phase of the Anthropic and AMD deal in 2027
- Whether other labs adopt Claude Mythos to move to Google TPUs
- Nvidia's response to the commoditization of AI hardware
Try it
Look up the current market share of Nvidia's AI compute contracts to see how much it has declined from its historical highs.
Sources and links
Read the original at Crypto Briefing
First reported by Crypto Briefing. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.