AI flywheels 2026-10-14

Reports

Notes

  • Performance flywheels in model, compute, data:
    • AI automated AI research
    • Revenue-compute flywheel: Inference revenue -> more access to frontier compute & R&D -> better and more cost-efficient models
    • Data flywheel: more high-quality usage and feedback -> better models & software. Strongest in coding and ads.
  • Economy of scale: higher compute utilization and lower R&D as share of revenue -> lower cost for unit of frontier intelligence.
  • Data and workflow lock-in
  • Performance flywheel mechanisms:
    1. Test-time compute
    2. RL
    3. Usage produces training signal -> product ownership matters.
    4. Distillation: performance created at the frontier becomes cost reduction one tier down, which also leaks.
    5. Speed is itself performance, especially for agents where latency compounds.
  • Cost flywheel mechanisms:
    1. Hardware generations, measured per watt.
    2. Model generations, on installed hardware.
    3. Model efficiency and routing: smaller distilled models, MoE designs, lower-precision maths.
    4. Custom silicons at lower prices.
    5. Disaggregation and specialization. e.g. Cerebras and AMD.
    6. Scale and utilization.
  • The compute economy
    • Unit economics driven by Jevons effect: The cost flywheel deflates the price of any given capability, while spend migrates to the frontier and to more volume (more tokens per task and new tasks, while frontier prices hold steady.
    • Risk of capability plateau: The key risk is a plateau in useful capability and the cost flywheel would turn into straightforward price deflation.
    • Risk of accelerated depreciation: if supply catches up with demand, older fleets reprise sharply.
    • Rising input costs work against the cost flywheel - HBM and memory.
    • Treadmill economics: The performance flywheel is a race to stay ahead of the cost flywheel's commoditization.
    • Improving gross margin and negative FCF: The cost flywheel helps gross margin. The performance race consumes it. Total economics are dominated by the next round of compute for training RL and capacity, so free cash flow stays deeply negative.
    • Compute and financing are the binding constraints. The suppliers of compute are underwriting the demand for compute, which is the clearest fragility in the system.
  • Strategic responses:
    • Own the product, not just the model. Coding agents, enterprise agent platforms, and consumer apps give a proprietary feedback loop (performance flywheel), customer stickiness, and pricing power.
    • Get your own chip. When computer is scarce only the chip supply is a competitive weapon not just a cost saving.
    • Multi-source compute. The labs are using the cost flywheel's competition among suppliers to bargain.
    • Distribute through every cloud for scale.
  • Company specific:
    • OpenAI: Hurt by lower-quality consumer usage data.
    • Anthropic: Own coding agents, high-quality coding usage data.
    • GOOG: Most complete flywheel owner with the model, the chip, the cloud and mass distribution. TPU external sales are a new way to monetise the cost advantage.
    • Nvidia: Compete on fungibility, co-design speed and financing. The cost flywheel is NVIDIA’s product roadmap, and the Jevons effect is its demand engine. The risks are custom silicon at the largest buyers and its growing balance-sheet exposure to labs.
    • AVGO, MRVL: Direct beneficiaries of inference specialisation. Watch customer concentration and memory-driven margin dilution.
    • AMZN, MSFT: Trainium and Maia are margin levers. The multi-model clouds (Bedrock, Foundry) monetise volume regardless of which lab wins. Data lock-in (customers want inference close to their data). Scale and utilization (global "fungible" fleet).
    • META: Owns chip (MTIA), model, distribution and usage data that give feedback (A/B test on steroids).
    • CBRS: A pure play on the speed tier, with heavy OpenAI concentration.
    • TSM, MU, SKHY, SNDK, ANET, CRDO, ALAB, COHR, LITE, CIEN: Suppliers whose volumes ride the Jevons effect. Memory suppliers are capturing part of the cost flywheel’s savings.
    • NET, AKAM: Edge inference is a possible beneficiary as smaller distilled models allow inference closer to users.
  • Five landscape shifts
    • From one token market to tiers: a commodity tier, a frontier tier, and a speed tier.
    • From general-purpose to specialized hardware. XPUs, co-designs, disaggregation.
    • From GPUs to full systems.
    • From token volume to token value. Open models handle a third of tokens but <4% of spend.
    • Suppliers becoming financiers.
Built with LogoFlowershow