SemiAnalysis: Global AI Computing Power May Be Dominated by Anthropic and OpenAI by 2028
Nashnova编辑部
Semiconductor research firm SemiAnalysis predicts Anthropic and OpenAI will control the majority of global AI compute by 2028, with generational leaps in chip efficiency as the key force driving that concentration.
How fast is new compute capacity growing?
SemiAnalysis founder Dylan Patel laid out a year-by-year path: ~30 GW this year, ~50 GW next year, ~70 GW the year after.
At that pace, by end of 2027 the two companies already account for roughly half of all new compute capacity.
This means → the global AI compute pie is nearly doubling each year, and half of every new slice goes to just two firms.
Why does "half" translate to dominance?
The key variable is a generational leap in chip efficiency — each new generation delivers 3 to 5 times the performance per watt of the last.
In plain terms = for the same power draw, a new chip does three to five times the work of an old one.
This means → even at only half of new capacity, the two firms' actual compute output would exceed the sum of all previously installed compute. Efficiency advantage stacked on share advantage produces an infrastructure monopoly.
Can this prediction hold up?
Patel's thesis requires two conditions to hold simultaneously: the two companies must keep locking in new data-center capacity, and chip efficiency gains must arrive on schedule.
If either falters — construction delays, process-node setbacks — the concentration timeline slows.
This reflects a deeper shift: the AI compute race is no longer about "who has more money" but about "who can lock down power, chips, and physical space at the same time."
Content is for reference only, not financial advice.