Eric Vishria of Benchmark says AI will produce an oligopoly of winners, not one company consuming everything, citing the big three clouds' roughly forty, thirty and twenty percent split.
Vishria says CEOs hitting their plans may be destroying equity value as public SaaS multiples fell from about thirty times revenue in 2021 to about six times.
Vishria says Cerebras's 2016 pitch that GPUs beat CPUs for deep learning was vindicated, yet hardware progress remains far harder than software because physics and supply chains set limits.
AWS history suggests AI will crown several winners, not one
The AWS era is a guide to AI. Early skeptics and later doomsayers were both wrong, because the market was undersized and produced a few large winners plus big companies outside them. Vishria applies that pattern to AI, where products get rebuilt every few months and SaaS economics are under pressure.
Takeaways
AI will produce an oligopoly, not one winner. Investors said in 2014 that AWS would eat all software, but Azure and Google Cloud were irrelevant then. Cloud now splits among three large players while Snowflake and Cloudflare grew huge outside them. Vishria expects AI layers to work while most companies fail.
AI products are sandcastles rebuilt constantly. Brett Taylor's metaphor fits Sierra, which builds around the jagged edge of models, as new capabilities arrive about every four weeks. Product managers must understand model capabilities, and taste matters most.
Agents make SaaS switching cheap and squeeze multiples. When agents translate databases, migration is cheap, so cost and elastic scaling decide winners. Public SaaS multiples fell from about thirty times revenue in 2021 to about six. CEOs hitting plans may be destroying equity value.
Inference is more specialized than it looks. Running very large open-source models is very hard. Fireworks delivers about five times the speed of a cloud provider running the same models, and that throughput advantage makes the business profitable on top of cloud margins.
Hardware bets can defy expert reasons against them. Five founders pitched Cerebras in 2016 on GPUs being a hundred times better than CPUs for deep learning. The wafer-scale chip had about 450,000 cores, and a working chip came around 2020, after the board saw heavy burn.
Chemistry and luck drive Benchmark's investing. Vishria is an investor second and a partner first, passing on deals without chemistry. He made eighteen investments in twelve years, and says getting many balls close to the pin increases luck.
Open problems Vishria is watching
Energy. China may add about ten times as much energy as the US next year, which could limit compute and intelligence.
Robotics. No internet-scale robot data exists, so high-quality pre-training data matters. Sunday Robotics uses laundry folding as a test task.
Radiology. Jeff Hinton's 2016 call to stop training radiologists was wrong on timing and data. New Lantern is working on it, and more radiologists will be needed as imaging demand grows.
Liquidity. Several hundred private SaaS companies between one hundred and five hundred million in size have no path to liquidity, and Vishria says public markets offer trust and capital.
Tap a name to go there · tap a line to see why they are linked