{
  "video": {
    "id": "laQ8wMysnkk",
    "title": "Buying GPUs Should Work Like Buying Stocks — Carmen Li, Compute Exchange",
    "duration": 1085,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "laQ8wMysnkk",
    "title": "Buying GPUs Should Work Like Buying Stocks — Carmen Li, Compute Exchange",
    "one_liner": "Compute Exchange is a neutral marketplace that owns no GPUs and lets you buy reserved and forward GPU contracts the way you buy stocks on an exchange. Carmen Li's Silicon Data indices show GPU prices behaving like a commodity: the H100 and B200 term-structure curves have shifted up and flattened, and the implied forward rates expect prices to rise over the next 36 months.",
    "summary": "Carmen Li compares GPU buying to stock trading. With 300+ neoclouds, token factories and API routers, calling 300 salespeople doesn't scale, and she argues buyers need a neutral, verified exchange that owns no GPUs, so it has no conflict of interest. Compute Exchange sells reserved GPU contracts, GPU and token forward contracts, and secondary/refurbished server sales. It verifies providers with benchmarks and checks buyers with KYB/KYC. Using Silicon Data (the 'Bloomberg for compute'), she shows H100 and B200 term-structure and implied-forward curves, plus residual-value curves. Her argument is that GPUs carry commodity-style term-structure risk that can be hedged with futures, swaps and block trades.",
    "key_points": [
      "Compute Exchange owns no GPUs by design. Li says owning GPUs would be like Nasdaq owning stocks: a conflict of interest that would let it cherry-pick clients and hoard relationships. It introduces buyers and providers to each other directly.",
      "Scale: 100+ verified providers, 80,000+ GPUs in supply, all regions except China, and a 24-hour turnaround commitment. It sees over a billion in RFQs a month, ranging from 1 node to 20,000 nodes.",
      "Verification is what separates it from brokers. For existing capacity it checks UUIDs, inventory and performance using Nvidia DGX, InferenceX and MLPerf. For forward capacity (e.g. GPUs starting Jan 1, 2027) it reviews purchase orders and permits. It runs KYB/KYC on buyers on behalf of neoclouds.",
      "Products: reserved GPU contracts (minimum one node for one month, up to 3–5 years, no on-demand); GPU forward contracts; and token forward contracts, a dedicated model deployment on reserved bare metal with fixed SLAs, quantization and precision, meant to fix volatile on-demand token API pricing. Buyers who over-commit on quantity or duration can sell capacity back.",
      "There has been a busy secondary server market for about the past 12 months, starting with requests like '20,000 H100 sitting at data center'. It covers everything from 4090s to B200s and AMD, and B200s sometimes trade above MSRP, which Li compares to used cars during COVID.",
      "Li draws on her energy-trading background: WTI is physically delivered and Brent is cash-settled. Compute Exchange's forward market aims to be 'the WTI' of compute and is a precursor to physically delivered futures. Silicon Data provides the indices that CME will use to settle futures on H100 neocloud on-demand rates and B200 prices, due to start trading within weeks.",
      "Term structure: last year the H100 reserved curve sloped downward (longer contracts were cheaper). Early this year it shifted up into contango, and it has since shifted further up and flattened. Implied forward rates, calculated with a Treasury-futures-style method, show a market consensus that prices rise over the next 36 months. B200 shows the same pattern.",
      "Silicon Data's residual-value curves use discounted cash flow, not the usual 5-year straight-line depreciation. NVIDIA's September investor day deck cites them, and several banks use Silicon Data to appraise GPU servers on their balance sheets. On those curves, A100 value declines but stays above straight-line depreciation, H100 value has risen, and B200 is above MSRP."
    ],
    "takeaways": [
      "When you need reserved compute soon, use a verified marketplace RFQ instead of calling dozens of provider sales teams. List several acceptable regions and state your interconnect, storage and software requirements, and compare offers on SLAs and benchmarks as well as price.",
      "Watch the reserved-contract term structure. In a flattened or contango market, signing longer terms no longer gets you a discount, so decide contract length with that in mind.",
      "If you hold GPU servers with 5–6 year financing but reserved contracts covering only the first few years, hedge the uncovered years (and counterparty default risk) with swaps or block trades priced against compute indices.",
      "For steady inference workloads, consider token forward contracts (dedicated deployments with fixed SLAs, quantization and precision) to avoid volatile on-demand API pricing.",
      "Don't value GPU assets with 5-year straight-line depreciation alone. Market-based residual values can differ a lot: the H100 has appreciated and the B200 trades above MSRP."
    ],
    "topics": [
      "gpu procurement",
      "compute markets",
      "futures and hedging",
      "term structure",
      "secondary hardware",
      "neoclouds",
      "inference pricing"
    ],
    "tools": [
      "Compute Exchange",
      "Silicon Data",
      "CME",
      "NYSE",
      "NASDAQ",
      "DRW Trading",
      "Nvidia DGX",
      "InferenceX",
      "MLPerf",
      "NVIDIA",
      "H100",
      "B200",
      "A100",
      "AMD"
    ],
    "quotes": [
      {
        "text": "Once we start owning GPUs, that would translate to just like Nasdaq owning stocks, translate to I have a conflict of interest with you and then with the providers.",
        "at": "02:29",
        "url": "https://www.youtube.com/watch?v=laQ8wMysnkk&t=149s"
      },
      {
        "text": "So we're creating that the WTI at Compute Exchange.",
        "at": "05:23",
        "url": "https://www.youtube.com/watch?v=laQ8wMysnkk&t=323s"
      },
      {
        "text": "Hey Carmen, I have like 20,000 H100 sitting at data center. Can you take away from me?",
        "at": "06:35",
        "url": "https://www.youtube.com/watch?v=laQ8wMysnkk&t=395s"
      },
      {
        "text": "GPU is behaves just like any commodity, have a very clear term structure risks, which data futures are excellent at hedging.",
        "at": "14:57",
        "url": "https://www.youtube.com/watch?v=laQ8wMysnkk&t=897s"
      }
    ],
    "words": 3226
  },
  "summary_url": "/#laQ8wMysnkk",
  "transcript": {
    "html": "/transcripts/laQ8wMysnkk.html",
    "txt": "/transcripts/laQ8wMysnkk.txt",
    "vtt": "/transcripts/laQ8wMysnkk.vtt"
  }
}