ALL INPredictions
Last updated Nov 29, 2025
Prediction
Friedberg
← Back to hostEpisode
Friedberg @ 01:01:58Inconclusive
techai
Over the next 1–2 years (by roughly 2027), hyperscale cloud providers like AWS, GCP, and Azure will add so much AI/GPU capacity and bundle it into their own services that third‑party GPU cloud providers like CoreWeave will face significantly reduced demand and downward pressure on pricing, similar to how "speed doubler" services became obsolete once broadband arrived.
That would be my biggest if I was to do diligence on this business. That's where I would spend a lot of my time is like, guys, what's the capacity going to be in a year or two? Sort of like when broadband hit the internet and you didn't need speed doublers anymore, do you really need to be paying as much as you are today? Is there going to be as much demand? How much is this going to get bundled in with GCP or AWS and so on in the future?View on YouTube
Explanation

The prediction was framed over a 1–2 year horizon from March 2025 (i.e., roughly until March 2027), so as of November 30, 2025 the full timeframe has not elapsed.

Evidence so far is mixed:

  • Hyperscalers are indeed massively expanding and bundling AI capacity. AWS and Google Cloud have rolled out successive generations of custom AI chips (Trainium / Inferentia, TPU v5e/v5p, v6 Trillium, v7 Ironwood) and large-scale GPU clusters, explicitly marketed as cheaper, integrated options for AI training and inference. (michaelbommarito.com)
  • GPU compute prices are already under downward pressure. Analyses of H100 rental markets in 2025 report a substantial increase in supply and intense competition, including a ~44% AWS price cut on H100 instances and broad price reductions across many clouds, with H100 rental rates drifting down across the market. (intuitionlabs.ai)
  • However, demand for specialized GPU clouds like CoreWeave has not “significantly reduced” so far; it remains very strong. CoreWeave has continued rapid expansion in 2025: it signed an ~US$12B, five‑year cloud deal with OpenAI, grew to 32 data centers with ~250,000 GPUs, became the first to offer Nvidia GB200 NVL72 (and later Blackwell Ultra) in the cloud, and completed a large IPO despite leverage and governance concerns. (en.wikipedia.org) Analyst commentary in mid‑2025 still highlights strong projected revenue growth into 2026–27 rather than collapsing demand. (barrons.com)
  • Other “neo‑cloud” GPU providers also show robust demand. For example, Nebius (another specialized AI cloud) reported several‑hundred‑percent revenue growth in 2025 and signed multi‑billion‑dollar, multi‑year infrastructure deals with major customers including Meta and Microsoft, noting demand was strong enough that one deal had to be capped by available capacity. (reuters.com)

So far, the observable market aligns with the pricing‑pressure part of the thesis (falling GPU rental prices as capacity ramps), and with hyperscalers increasingly bundling AI compute into their platforms. But the “significantly reduced demand for third‑party GPU clouds” component has not materialized yet; if anything, those providers are still rapidly scaling and signing large contracts. Because the prediction’s endpoint is still ~16 months away and current data do not decisively confirm or refute a future demand drop by 2027, the outcome must be judged inconclusive (too early) rather than clearly right or wrong at this time.