Baker proposes a new metric decomposing MFU that he developed the morning of the interview.
“MFU is the most important metric because it gives you all of these advantages and ways to differentiate yourself amongst five people, six people who've trained these GPT core class models.”
Baker explains higher MFU allows 25% faster time to market with same GPU and power spending.
“You have the same amount of GPUs and the same amount of power presumably. You could choose between faster time to market.”
Baker explains MFU runs at 35-40%, measuring the percentage of theoretical compute actually used for training.
“MFU, model flops utilization, and that generally runs around 35 to 40%. And that's literally the percentage of compute, theoretical compute flops that you're actually applying to trading.”
Baker explains that 50% MFU versus 40% MFU enables 25% faster time to market for AI models.
“You could choose between faster time to market. If you run a 50% MFU and your competitor's running 40 for an equivalent amount of trading flops, you could be in market 25% faster.”
Memory bottlenecks create severe performance penalties, with a thirty-eight times slowdown when data operations exceed available DRAM capacity.
““38x slowdown once shuffle data exceeds available DRAM and spills to SSD.” From a joint Micron/Meta white paper. The DRAM bottleneck almost certainly lasts the longest. OP from @jukan05”
“Market is overreacting to hyperscale credit spreads widening from my perspective. TL;DR Spot pricing for renting GPU compute materially above contracted rates implies hyperscalers are underearning while operating cash flow acceleration is an underestimated source of funds for AI https://t.co/D9FSHXQ8F8”
“@NoonHillCapMgmt Always wise to be skeptical. Nothing wrong with the analogy, just the barriers to entry are higher, quality of companies is materially higher and structurally harder to bring on capacity given shortages.”
“A series of fairly wild thoughts from Dwarkesh. I thought I was bullish but even I am not assuming that the price to rent compute continues to go up! I will say that I am in SV this week and Dwarkesh is capturing the zeitgeist. Net, net public markets would probably be trading https://t.co/PzIxxnzbiS”
“@jukan05 4 gigawatts this year and another 4 next year per this report. Spot price for compute has gone up since they signed their Anthropic and Google deals.”
An observer interprets Baker as implying the AI compute business could reach one hundred billion in annual revenue near term.
“If you're a $SPCX shareholder take notes 📝 Gavin Baker, an OG investor just said in VC terms that the AI compute business is about to get WAY BIGGER than just the Google & Anthropic deals He's implying like $100B in rev/yr on the table near term if I'm reading it right”
Baker predicts vertically focused AI native companies will accelerate due to routers, open-source models and specialized post-training
“We are going to see a lot of vertically focused AI native companies accelerate. Routers, open-source models and specialized post-training enabled by companies like @FireworksAI_HQ have all made dramatic advances and the combination of the three is driving accelerating growth. https://t.co/f8qJdKSciI”
Baker argues game theory suggests memory LTAs are durable and Nvidia may have significant cloud revenue shares
“Game theory suggests memory LTAs are likely to be durable. And yes, Nvidia may have a significant cloud business via revenue shares. https://t.co/GFN1kDQmSk”
Baker describes a playbook where companies prove strong returns on capacity then raise capital to scale.
“ watched analysts reward it, and is widely speculated to be planning an equity raise. The playbook: prove strong IRRs on a slice of capacity, raise capital, then ”
Baker claims installed compute capacity trades at massive discount to spot prices making repricing wave inevitable.
“ are still going vertical in 2026. And because the entire contracted base of installed compute — every long-term deal signed over the past two years — trades at a massive discount to current spot prices, a repricing wave is inevitable. ”
Baker identifies memory as the single most important factor for increasing token output per compute unit.
“It's the single most important thing you could do to increase token output per unit of compute, and then that obviously, definitionally, actually lowers costs,”
Baker suggests NVIDIA accelerators have residual value that distinguishes them from first-party ASICs, responding to concerns about financing OpenAI hardware
“@dylan522p @TShirtnJeans2 @maxkan Some accelerators (the green ones) are not like the other accelerators.”
Baker questions whether AFD is compatible with data locality, disputes OpenAI bet against disaggregation
“@maxkan Hmm. I might ask @dylan522p if Attention-FFN disaggregation (AFD) is compatible with data locality and not moving the KV cache. That constraint is what OpenAI focused on with Jalapeño. They absolutely did not make an “explicit bet that disagg is not the way to go.” There are https://t.co/2lVjH2M5vS”
Baker hypothesizes tomorrow's Ultrafast will be GPUs or Jalapeño with CS-4 or CS-5
“@dylan522p @maxkan And they also said this followed by “Cerebras FTW.” I would be surprised if their head of Codex was unaware of their roadmap. Reasonable to hypothesize tomorrow’s Ultrafast will be either GPUs or Jalapeño working with CS-4 or CS-5. We shall see! https://t.co/QM0vs809x7”
Baker speculates Ultrafast will use GPUs plus CS-4/5, notes Jalapeño not financeable at GPU rates
“@dylan522p @maxkan Maybe GPUs plus CS-4/5 for Ultrafast at scale. Jalapeño not financeable at comparable rates to GPUs and probably not that big of a wafer allocation next 2 years. Jalapeño still a really impressive achievement from OpenAI. I remember being sad when they hired Thomas away from https://t.co/a0jIXXahXN”
Baker endorses data centers as beneficial for America, reacting to Quincy Washington's economic transformation
“Data centers are awesome for America in every way. https://t.co/gSe3Gf2kYp”