A look at the Nvidia A100, a ~$10K GPU that has become a critical tool for generative AI; Nvidia has an estimated 95% market share of machine learning GPUs
Kif Leswing / CNBC : Tweets: @nightingale_p , @stagefright__ , and @sarbjeetjohal Tweets: Paul Nightingale / @nightingale_p : AI infrastructure isn't cheap. Making chips isn't easy. $80bn on upgrading servers a chunk of change for anyone.... https://twitter.com/... Ilija / @stagefright__ : 8 high-powered graphic cards each at 300W to compute a single answer to your chatGPT quip. if true that's at least 2400 Watts thrown at every question. for context my Macbook's (M1 Max) chip draws around 65W at max load https://twitter.com/... https://twitter.com/... Sarbjeet Johal / @sarbjeetjohal : .@nvidia jumps 14% after stellar earnings and a rosy outlook. Around 60% of their revenue comes from datacenter, most of that is from #Cloud #hyperscalers. IMO Nvidia will keep benefitting from rise of #AI #MWC23 #Barcelona @dvellante @furrier https://www.cnbc.com/...
Context & Ripple Effects
The A100 profile lands mid-arc for a company that spent a decade positioning for exactly this moment: after missing out on smartphone silicon, Nvidia rebuilt itself around AI, self-driving, and datacenter demand, and by early 2023 its bet is paying off as generative AI makes the ~$10,000 GPU the scarce input everyone needs.
The economics in the piece are the story: an estimated 95% share of machine learning GPUs, roughly 60% of Nvidia's revenue from datacenter sales to cloud hyperscalers, and a single ChatGPT-style answer potentially drawing kilowatts of compute. Weeks later, CEO Jensen Huang walked through the H100's transformer engine and the US-China export tension on camera — evidence that both demand and geopolitics were already outrunning supply.
First-order effects
- Cloud hyperscalers face a straight capex bill — the cited $80bn in server upgrades flows largely into Nvidia's order book at ~$10K per A100, giving the company near-monopoly pricing power over anyone building generative AI services.
- Every AI operator now carries a visible per-query energy cost: the reported figure of eight high-powered cards drawing ~2,400W per answer means inference efficiency, not just training scale, shows up directly on the power bill.
Second-order effects
- Intel and AMD are left scrapping for second place in datacenter GPUs — a gap later measured at Nvidia holding 90%+ share even after its Q1 data center revenue grew 427% YoY to $22.6B (IoT Analytics via Nikkei Asia) — while Nvidia keeps refreshing the top of the market with platforms like the DGX GH200 supercomputer.
- Sovereign buyers enter the queue alongside the hyperscalers: Saudi Arabia's purchase of 3,000+ H100s and the UAE's thousands of GPUs (per the Financial Times) show national AI programs competing for the same constrained supply, tightening allocation further.
Third-order effects
- If one vendor holds ~95% of the machine-learning GPU market, AI capacity starts behaving like utility infrastructure — metered, capital-intensive, and strategically regulated — which invites both export controls and buyer diversification efforts as structural responses rather than one-off news events.
- The per-query wattage gap between datacenter GPUs and consumer chips pushes the industry toward treating compute cost as a first-class design constraint, favoring players who control the full stack from silicon through software.
The trend: Generative AI is turning GPU supply into utility-scale infrastructure, concentrating pricing and geopolitical leverage in a single vendor's product roadmap.