"The functional reason for commodity it is for hedging right... really for people to hedge their volatility to do risk allocation to do risk transfer and then asset capital allocation if we can't do what you said then we fail at our job." - Carmen Li [00:00:50]
"You can't you don't know until you get your GPUs... we proved there's 38% performance variance for the same chip and then we decomposes into the chip itself intra provider and interprovider." - Carmen Li [00:09:28]
Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer
"Our index model is not a simple math. It's not hey you have two H100s do simple average right because then you compare apple to oranges." - Carmen Li [00:11:53]
"Last year when GPU prices all going down the big conversation is why do you need indices for something price will always go down and this year is why do you want indices when price always go up." - Carmen Li [00:13:27]
"The second year H100 residual value resale value for refurbish chips about 85 cents on a dollar so a year later you can sell 85 cents $1 that's pretty good... the third year is 84 cents $1 and my car depreciate way more than that." - Carmen Li [00:29:01]
"I look at bubble is can your future cash flow support today's valuation of yours right... look at future cash flow of your forward contracts and then you discount it back." - Carmen Li [00:30:40]
Speakers & Credentials
Tracy Alloway: Co-host of the Odd Lots podcast by Bloomberg. Veteran financial journalist covering market structures, derivatives, and systemic risk.
Joe Weisenthal: Co-host of the Odd Lots podcast by Bloomberg. Expert in macroeconomic trends, financial history, and emerging asset classes.
Carmen Li: CEO of Compute Exchange and Silicon Data. Pioneer in creating financial indices for computational power, architecting the first global GPU indices on the Bloomberg terminal, and spearheading a partnership with the CME Group to launch compute-based futures and options.
1. Executive Summary
Compute power, particularly GPU capacity, is rapidly transitioning from a localized physical asset into a globally financialized commodity, mirroring the historical maturation of the crude oil market.
The inherent fungibility problem of hardware is being solved through complex index-weighting mechanisms that track 8 million global pricing points to create standardized tradable instruments.
Because individual hardware exhibits up to a 38% performance variance for the exact same chip model, standardized contracts and independent "carfax" style verifications are necessary to clear the market.
Compute Exchange and Silicon Data, in partnership with the CME, are launching financially-settled compute futures (pending CFTC approval) to allow NeoClouds, AI startups, and banks to hedge their natural long or short positions.
Despite prevailing narratives of an imminent AI bubble, secondary residual values of high-end chips remain robust, with H100s retaining 85% of their value in year two, indicating sustained foundational demand.
Volatility in the underlying spot market is surprisingly healthy for a financial instrument, averaging 20-30% daily variance, giving speculators and hedgers alike ample economic rationale to utilize standardized derivatives.
2. Chronological Table of Contents
[00:00:00] Introduction: The Core Philosophy of Hedging Compute
[00:01:51] The Analogy: Why Compute is the New Oil
[00:04:24] Launching CME Futures & Bloomberg Indices
[00:05:52] Spot Market Dynamics & The Shift to Forward Contracts
[00:09:04] The GPU Lottery: Solving for 38% Performance Variance
[00:11:53] Constructing the Index: 150,000 Daily Price Ingestions
[00:15:22] Sourcing the Data: 8 Million Pricing Points
[00:24:50] Recent Pricing Trends: H100, B200, and A100 Trajectories
[00:29:01] The "Carvana Model": Refurbished Chips & 85% Residual Value
[00:30:40] The AI Bubble Question & Future Cash Flow Valuations
3. Detailed Thematic Summary
The Financialization of Compute & The Forward Market Shift
The compute market is mirroring traditional commodities, offering on-demand, reserved, and forward contracts [00:06:38].
On-demand prices offer flexibility but expose users to wild swings, shifting from $3 to $6 to $9 per hour based on supply and demand curves [00:06:52].
To stabilize predictable margins and secure scarce resources, enterprises and AI startups are heavily pivoting to forward contracts, locking in deliverables months in advance [00:07:13].
The introduction of financially settled CME futures will allow entities "naturally short" compute (most AI developers) and entities "naturally long" (NeoClouds and Banks holding hardware on balance sheets) to properly hedge their exposures [00:08:21].
These CME contracts will be traded utilizing standard prime brokers with familiar margin optimization protocols, rather than requiring exotic new trading infrastructure [00:21:40].
Overcoming the Fungibility Barrier: Indexing and The "GPU Lottery"
Unlike digital assets or standardized gold bars, compute is not homogeneous; a straight mathematical average of two H100 chips is useless due to variances in RAM, disk speed, location, and CPU configurations [00:11:53].
Silicon Data and Jefferson Lab published a paper proving a 38% performance variance for the exact same A100 (40 GB memory bandwidth) chip depending on deployment [00:09:28].
To solve this "GPU Lottery," the platform acts as a "GPU Carfax," independently verifying flops, memory bandwidth, and tokens before delivery, allowing buyers to price-in geolocation and latency premiums [00:09:52].
The index model normalizes data by ingesting 150,000 traded prices daily across 100+ sources, establishing a base case that isolates the variables driving price differentiation [00:12:22].
This massive data collection expands to 8 million global pricing points aggregated from 200 different data sources to form an accurate, manipulation-resistant spot benchmark [00:15:22].
The compute market is intentionally being structured to emulate the global crude oil market, specifically utilizing the WTI and Brent indices as foundational mental models for base pricing [00:02:38].
In this paradigm, NeoCloud operators and hyperscalers act identically to shale oil producers—they are the natural holders of physical reserves that must short futures to hedge against revenue collapses [00:07:50].
Similar to agricultural and energy commodities, the compute market exhibits extreme local variances; some localized chip deployments display 80% to over 100% volatility, though index normalization stabilizes daily volatility tracking to a normalized 20% to 30% [00:14:03].
When addressing the macroeconomic fears of an "AI Bubble," Carmen explicitly invokes the historical context of the dot-com Nasdaq crash, where the index shot up 200% before collapsing 84% [00:30:40]. However, she differentiates hardware economics by anchoring value strictly to the discounted cash flow of forward leasing contracts rather than speculative equity multiples.
Price Inflexions, Residual Value & The Token Market
Despite being older generation technology, A100 lease prices have unexpectedly increased by 10-15% over the past three months, signaling a massive structural supply shortage that has yet to clear [00:26:12].
Similarly, H100 rental rates on-demand have increased approximately 8% in the last 90 days [00:26:03].
The LLM token index (a separate metric measuring inference costs rather than raw compute) has doubled since December, reaching $2.21 per million tokens, driven by consumption-weighted averages of massive scale models [00:24:14].
Addressing concerns regarding chip lifespan depreciation (sparked by Michael Burry's tweets), data reveals that refurbished H100s maintain immense secondary value, retaining 85 cents on the dollar in year two, and 84 cents on the dollar in year three [00:29:01].
Even legacy "OG" chips like L40s are still commanding 40 cents per GPU per hour from hyperscalers, proving the long-tail economic viability of older silicon [00:29:45].
The Reference Vault
4. Data & Figures
Data Point
Value
Context
Timestamp
A100 Performance Variance
38%
The documented performance difference for the exact same A100 chip based on inter/intra provider discrepancies.
The Brent/WTI Energy Blueprint for Compute
By mapping the emergent AI ecosystem directly onto the mid-century structural evolution of crude oil markets, analysts can bypass the noise of tech valuations. In this model, physical data centers and hyperscalers represent the capital-intensive upstream drillers (long physical assets, vulnerable to price crashes), while software layers and LLM developers represent the downstream refineries and consumers (short physical assets, vulnerable to price spikes). This framework proves that compute futures are not a speculative casino, but rather the fundamental plumbing required for macroeconomic risk transfer and structural capital allocation. [00:02:38]
The GPU Lottery & The Fungibility Paradox
A defining structural barrier to the commoditization of technology is physical deployment divergence. Two identical Nvidia chips will not perform identically; network topography, cooling efficiencies, memory bandwidth, and intra-provider routing create massive variances (up to 38%). Therefore, the framework for trading compute cannot mirror digital currencies or pure energy. It requires an indexing framework analogous to basis risk in agriculture, where a core normalized baseline is established, but participants must calculate geographic and performance "basis points" to clear physical delivery or hedge localized edge deployments. [00:09:28]
The "Carvana Model" of Silicon Residuals
Traditional hardware depreciation assumes a rapid plunge toward zero as Moore's Law renders obsolete older silicon. The new mental model dictates a hyper-resilient secondary market, driven by intense capital rationing. Just as used car prices stabilized at historical highs due to supply chain shortages, older generation GPUs (like the L40s and A100s) maintain immense baseline cash-flow capabilities. Evaluating GPU ROI now relies on calculating a slow-decaying residual asset (84 cents on the dollar by year three) rather than writing the asset off, fundamentally shifting how debt financing for data centers is modeled. [00:29:01]
The Fundamental Hardware Valuation Model (DCF over VC Hype)
To answer the "AI Bubble" debate, one must conceptually separate the software equity market from the infrastructure commodity market. While software startups are valued on speculative multiples of eventual market dominance, the underlying hardware is valued using a strict, traditional Discounted Cash Flow (DCF). By analyzing the guaranteed revenues of locked multi-year forward contracts against the residual physical value of the servers, one can mathematically isolate whether the physical build-out of AI infrastructure is a speculative bubble or a mathematically sound infrastructure play. [00:30:40]
6. Anecdotes
The Polymarket Rogue ListingContext & Purpose: To illustrate the organic, pent-up demand from the speculative market to trade compute volatility, Carmen shared how her indices were hijacked. Before securing formal institutional partnerships, a decentralized prediction market (Polymarket) listed her proprietary compute index without her consent. Rather than purely fighting it, she used it as a sandbox, structuring formal February and April settled contracts to test liquidity waters. It proves the cultural and economic inevitability of compute derivatives; if institutions didn't build it, decentralized crypto actors would have forced it. [00:22:21]
The 6-Month Contrarian Reality CheckContext & Purpose: To demonstrate the sheer unpredictability of spot compute pricing, Carmen reminded the hosts of their previous conversation. Six months prior, prevailing market consensus and journalistic questioning revolved entirely around falling compute prices as supply supposedly met demand. Today, those identical questions must address why older, supposedly outdated chips (A100s) are appreciating 10-15%. It serves as an anecdote on why rigid linear forecasting fails in exponential tech rollouts, cementing the urgent need for a hedging instrument. [00:03:58]
The Michael Burry Obsession with Chip DepreciationContext & Purpose: During the height of macro-anxiety surrounding AI investments, famous short-seller Michael Burry tweeted about the potentially rapid degradation and short lifespans of AI hardware. The financial media engaged in a three-week frenzy of panic regarding chip lifespans. Carmen utilized this anecdote to highlight the disconnect between financial Twitter narratives and hard physical data. By simply running the residual pricing code, she proved the hardware depreciates slower than a standard commuter vehicle, neutralizing a viral, highly speculative narrative with cold basis data. [00:28:18]
7. References & Recommendations
Geopolitical & Financial Institutions
CME (Chicago Mercantile Exchange): The primary exchange partnering to list the financially settled compute futures, bringing traditional margin requirements to tech hardware. [00:04:24]
CFTC (Commodity Futures Trading Commission): Mentioned as the regulatory gatekeeper pending approval for the launch of compute futures. [00:04:24]
Companies & Entities
Silicon Data & Compute Exchange: The primary platforms discussed, acting as the index provider and the spot/forward marketplace respectively. [00:03:15]
Bloomberg: The financial data terminal utilized to launch the world's first global GPU index and currently tracking the LLM token prices. [00:04:48]
Jefferson Lab: Co-authored the core research paper confirming the 38% variance in hardware performance due to deployment variances. [00:09:04]
Polymarket: The decentralized prediction market that organically listed unauthorized, then authorized, compute derivatives. [00:22:21]
DRW: A high-frequency trading firm led by Don Wilson, heavily referenced as the primary capital backer of the compute exchange architecture. [00:01:37]
People
Don Wilson: Founder of DRW, an initial catalyst and investor driving the structural financialization of computing power. [00:01:37]
Michael Burry: Famous investor who sparked a short-lived panic regarding the rapid depreciation and physical lifespan limits of AI silicon. [00:28:18]
Historical Events & Concepts
The Dot-Com/NASDAQ Crash: Specifically referenced the historical 200% climb and 84% crash to define what a purely speculative financial bubble actually looks like in contrast to hardware cash flows. [00:30:40]
Commodities & Hardware Architecture
Brent & WTI Crude: The global benchmark grades for crude oil, utilized entirely as the template for creating a centralized compute index. [00:02:38]
Nvidia H100, A100, B200, L40s: Specific generations of GPUs whose disparate residual values, volatility arcs, and lease rates are the foundational assets being tracked. [00:24:50]
8. The Bottomline (by AI)
The transition of computing power from a localized tech expense to a fully financialized global commodity marks a paradigm shift in how capital expenditures are modeled for the next decade. Standardized compute futures will unlock massive institutional liquidity by allowing hyperscalers to secure financing against hedged hardware yields, effectively establishing the "Brent Crude" equivalent of the 21st-century digital economy. Market participants must immediately recalibrate their CapEx models to account for a hyper-resilient secondary hardware market, shifting focus away from software valuation multiples and toward the structural basis risk of physically delivering scalable compute.
Jul 16, 2026
Dr. Robert Wachter | A Giant Leap: How AI Is Transforming Healthcare... | 14 Jul 2026 | Talks at Google
"don't get me wrong US healthcare delivers miracles every day particularly when it comes to cutting edge and intensive care... but the health care system itself is a headache wrapped in red tape inside the nightmare that France Kofka himse…
Global Pricing Dataset
8 Million
The total volume of pricing data points collected globally from around 200 data sources.