AMD and the Hard Business of Challenging a Software Moat

The accelerator market looks like it should be competitive. Demand is enormous, margins are extraordinary, and every large customer is desperate for a second source. Yet the leader’s share has proved remarkably durable.

The reason is not primarily silicon. It is the accumulated software ecosystem that developers have built on top of one company’s programming model over more than fifteen years.

What the incumbent actually owns

• Libraries that researchers depend on, many written by third parties for one platform only.
• Kernel implementations optimised for one architecture’s memory hierarchy.
• A hiring market where every machine learning engineer already knows one toolchain.
• Institutional inertia: rewriting a training pipeline is expensive and risks regression.

Make porting cheap rather than win on raw performance

The credible approach is not to beat the incumbent at its own game but to make porting cheap. That means investing heavily in translation layers, funding open-source libraries to add support, publishing optimised kernels for the workloads that matter, and — most importantly — being present in the places where new projects start.

Open-weight models have helped considerably here, because they let a challenger demonstrate performance on real workloads without needing a frontier lab’s cooperation.

A second source for inference, not a new leader

Nobody serious expects the incumbent to be displaced. The plausible outcome is a genuine second source capturing a meaningful share of inference — where workloads are more standardised and cost sensitivity is highest — while the hardest training runs remain on the established platform for some years.

That would still be a substantial business, and it would change the negotiating dynamics for every buyer in the industry.

Serving is the volume, and the opening

The most plausible foothold is inference rather than training. Inference workloads are more varied but individually simpler, they are extremely cost-sensitive, and many organisations are willing to accept a modest performance penalty in exchange for a substantial price advantage.

Inference also represents the majority of compute consumed over a model’s lifetime. A model is trained once and served many millions of times, so a small efficiency gain in serving compounds across the industry.

The libraries decide it, and they take years

Community support matters more to a challenger than to an incumbent. If widely used open-source libraries add first-class support for an alternative platform, porting becomes routine rather than a project, and the moat weakens through attrition.

That is a slow process and it depends on the challenger funding work that benefits competitors as much as itself. It is also the only route that has ever worked against an entrenched software ecosystem.