Comparing Nvidia's new A100 Ampere chip with competition from Intel and GraphCore in terms of performance, economics, and software
Now that the dust from Nvidia's unveiling of its new Ampere AI chip has settled, let's take a look at the AI chip market behind the scenes and away from the spotlight
Context & Ripple Effects
Days after Nvidia's A100 unveiling — 54B transistors and 5 petaflops, roughly 20x Volta — ZDNet steps back from the spec-sheet spotlight to ask how the chip actually stacks up against Intel and GraphCore on performance, unit economics, and software. That framing matters because the launch coverage was all peak numbers, while buyers choose on cost-per-workload and tooling.
The retrospective record vindicates the question: the A100 went on to become the ~$10K workhorse of generative AI with an estimated 95% machine-learning GPU share, while a later roundup of Nvidia's rivals shows AMD, Intel, startups, and cloud incumbents still chasing from behind.
First-order effects
- AI infrastructure buyers get a decision framework beyond raw petaflops: the comparison forces Intel's and GraphCore's accelerators to be judged on total cost and software maturity against a chip whose successor-generation leap reset expectations overnight.
- Intel and GraphCore are put on the defensive immediately — the A100's 20x generational jump makes their current parts look like point solutions rather than platform alternatives.
Second-order effects
- Competitors' responses shift toward the software layer, since matching transistor counts is easier than matching the CUDA-style developer lock-in the article singles out as part of the comparison.
- Cloud providers and hyperscalers gain negotiating leverage: credible second sources on paper give them pricing arguments even if Nvidia's installed base keeps orders flowing its way.
Third-order effects
- If the pattern holds, AI compute consolidates around one vendor's integrated hardware-plus-software stack, leaving rivals to compete for second source — a structure later data points confirm, with IoT Analytics putting Nvidia at 90%+ of data center GPUs while Intel and AMD scrap for runner-up.
- The durability of that concentration becomes a regulatory and procurement question, since buyers dependent on a single supplier's roadmap have limited fallback when capacity tightens.
The trend: AI accelerator markets are consolidating around Nvidia's integrated silicon-and-software stack, with rivals competing on economics and openness rather than raw chip specs.