# Why Compute Still Trades in Group Chats — Affaan Mustafa & Alejandro Soumah Fernandez, Jump Trading

Affaan Mustafa, Jump Trading · Alejandro Soumah Fernandez, Jump Trading

[Talk page](https://ai.engineer/talks/GCse4OMzY8U-why-compute-still-trades-in-group-chats) · [Video](https://www.youtube.com/watch?v=GCse4OMzY8U)

## Why Compute Still Trades in Group Chats

Affaan Mustafa and Alejandro Soumah Fernandez explain how physical infrastructure, startup financing and bespoke contracts keep GPU capacity trading privately—and what a more useful compute market would need to solve.

![The opening slide compares an equity dark pool with the private quotes, brokers and bilateral contracts used to trade compute.](https://ai.engineer/api/talk-essay-media/c324e708366f91e44809d5dcc4fcf66f9a8d3bc43b048b0bfe7e01504baa206e)

The opening slide compares an equity dark pool with the private quotes, brokers and bilateral contracts used to trade compute.

### A GPU quote is the beginning of a commercial agreement

[0:12](https://www.youtube.com/watch?v=GCse4OMzY8U&t=12s)

A company buying a fixed block of GPU capacity for several years may negotiate the deal in Slack messages, X messages or Telegram group chats. Public marketplaces exist, but much of this longer-term business happens through private quotes and bilateral agreements. Affaan Mustafa and Alejandro Soumah Fernandez, compute-trading co-leads at Jump Trading, begin with that gap between visible prices and the way substantial capacity commitments actually change hands.

Mustafa calls this an “unlit market.” His use of dark pools describes negotiations outside public view, rather than establishing that compute trades through the same machinery as a financial dark pool. The underlying transactions are commercial agreements: some combination of a master services agreement, a service-level agreement and an order form, with language specific to the cluster and its hardware.

That specificity matters because pricing conversations often collapse the product into a chip identifier: H100, H200 or another SKU. The actual purchase includes a particular amount of capacity, a term, a total contract value and cluster-specific obligations. An on-demand price reference can help start the conversation, but it does not describe the entire agreement being negotiated. There is no widely adopted physical-compute exchange that brings all those transactions into one public venue.

### Cheaper intelligence can require more hardware

[2:44](https://www.youtube.com/watch?v=GCse4OMzY8U&t=164s)

Falling token prices seem, at first, to imply falling demand for expensive GPUs. The missing variable is how much work people choose to run. Mustafa introduces a capability-cost graph attributed to a16z: as a given capability becomes cheaper, a team can do more within the same budget. Previously uneconomic applications become possible. The price of intelligence, the number of tasks requested and the hardware needed to serve those tasks are three different quantities.

How can token prices keep falling while compute prices hold up? From July through August 2026, the Silicon Data token price expenditure index fell while Ornn’s H100 rental price held relatively steady. The chart measures H100 rental in dollars per GPU-hour on the left axis and the token index in points on the right. Mustafa offers Jevons paradox as one possible explanation: more efficient inference reduces the resources needed per task, while cheaper tasks encourage enough additional use to outweigh that saving. The chart shows the price divergence but does not measure the change in usage.

![Two price series diverge on the slide comparing token and GPU capacity prices. The speakers identify workload mix and usable cluster supply as factors.](https://ai.engineer/api/talk-essay-media/71a36b8e3b430f15d0bd0a639926e823b8c1d2503f6f4cab630a061e438c8ffd)

Two price series diverge on the slide comparing token and GPU capacity prices. The speakers identify workload mix and usable cluster supply as factors.

The causal steps are straightforward. More efficient GPUs and serving systems lower token costs; teams spend their budgets on more tokens; inference clouds serve more applications; application developers can reach more customers. The capacity supporting all that activity still has to be built and paid for. A cheaper unit of output therefore does not, by itself, establish that the aggregate hardware bill should fall.

### What must come together before a GPU hour exists

[5:08](https://www.youtube.com/watch?v=GCse4OMzY8U&t=308s)

A chip is only one input to usable compute. Power, buildings, grid connections and capital must arrive together before a customer can consume a GPU hour. Mustafa cites an IEA April update projecting global data-center electricity use rising from 485 terawatt-hours in 2025 to 950 in 2030. The projection frames the practical questions facing operators and lenders: where to obtain land, how to secure electricity and whether the grid can support the planned build-out.

![The supply diagram connects chip and system vendors, site operators, capacity providers and compute users. Lenders and brokers appear below the chain.](https://ai.engineer/api/talk-essay-media/07b0cb628dacf5f7d5d55159025319c0bcf4355ea1869237c41acb12de1209cc)

The supply diagram connects chip and system vendors, site operators, capacity providers and compute users. Lenders and brokers appear below the chain.

Which activities converge to produce capacity, and who consumes it? The supply map separates equipment from the facilities that make it usable. Chip and system vendors supply GPU servers, memory and networking. Sites and operators supply power, space, assembly and management. Capacity providers—including hyperscalers, neoclouds and other bare-metal operators—make the resulting compute available to users. Upstream constraints feed into both its cost and its availability.

These are roles, rather than mutually exclusive kinds of company. A business can build the infrastructure, operate it and consume the compute itself. At the user end, the workloads span startups, inference clouds, labs, application companies and enterprises running open-weight models. The map explains the dependencies without requiring every GPU hour to pass through a separate company at each stage.

Parallel inputs converge into usable capacity
Equipment and operating infrastructure support capacity providers, which supply users. These roles can coexist inside one company; the arrows describe supply dependencies.
- Chip and system vendors — GPU servers, memory and networking
- Sites and operators — Power, space, assembly and management
- Capital — Finances the physical build-out
- Capacity providers — Hyperscalers, neoclouds and other bare-metal operators
- Compute users — Startups, inference clouds, labs, application companies and enterprises
- Chip and system vendors → Capacity providers (Supplies hardware)
- Sites and operators → Capacity providers (Supplies operating infrastructure)
- Capital → Capacity providers (Funds build-out)
- Capacity providers → Compute users (Supplies compute capacity)

### A prepayment can consume the funding round

[7:18](https://www.youtube.com/watch?v=GCse4OMzY8U&t=438s)

An inference cloud can rent a cluster, sell tokens from it and use that revenue to pay for the cluster. The difficulty is financing the commitment before the revenue arrives. Hyperscalers, frontier labs and larger neoclouds can raise substantial equity, take on debt and sign long-horizon agreements. A Series B startup facing a compute contract worth more than $200 million over three years has a much harder underwriting problem.

Lenders want confidence that a build-out will generate enough revenue to repay their loans. A young private company may have limited cash flow or no revenue at all, so contracted customer revenue becomes especially important. Mustafa points to CoreWeave's public filings as a large-scale example of infrastructure financing supported by take-or-pay customer contracts. Long commitments help support financing, while customers may be reluctant or unable to commit for 36 months.

The upfront cash requirement makes this mismatch concrete. Mustafa attributes a 15–25% prepayment range to CoreWeave; on a $100 million contract, that means $15–25 million paid before the startup can use the money elsewhere. He then switches to a separate illustrative case: a company raises $40 million, signs a $100 million compute contract over three years and prepays 30%. The $30 million prepayment leaves $10 million to run the business.

How does that financing constraint turn into a capacity constraint? The example describes contracted capacity growing as a step function: each new cluster requires another large commitment. Demand can sit below the purchased block, leaving capacity unused, then exceed the available capacity before the company can finance the next block. Growth can become tied to the next equity round, even when customers already want the service.

![In the slide’s example, a $40 million raise leaves $10 million for operations after a $30 million prepayment. A chart shows capacity commitments increasing in blocks.](https://ai.engineer/api/talk-essay-media/e73a16b8a42ac21185f2a34b7cbdf4e77c8abff9cfe8567e0126a5f07d3c9448)

In the slide’s example, a $40 million raise leaves $10 million for operations after a $30 million prepayment. A chart shows capacity commitments increasing in blocks.

The company has several responses, each with a different cost:

- **Turn customers away:** Wait until another financing round supports the next capacity commitment, sacrificing business in the meantime.
- **Bridge with on-demand capacity:** Buy enough short-term compute to cover the shortfall, provided that capacity is available and its price works for the business.
- **Finance the prepayment:** Borrow the upfront amount and compare that financing cost with the cost of covering the same need on demand.

The recording leaves this as an economic comparison rather than supplying a rate at which borrowing becomes preferable. Availability matters alongside price: an attractive on-demand quote cannot bridge a shortage if the required capacity cannot actually be obtained.

### Build, colocate or buy capacity

[10:36](https://www.youtube.com/watch?v=GCse4OMzY8U&t=636s)

Procurement also determines how much of the physical system a company controls. Mustafa lays out three operating models:

- **Build and operate:** Own the hardware and site, arrange power and permits, and handle assembly and management.
- **Own servers in colocation:** Retain server ownership while a partner handles more of the facility, assembly and operating work.
- **Sign an offtake contract:** Buy defined capacity from a provider that owns and operates the assets. Access can vary, including restrictions on root access and differences between single-tenant and multi-tenant configurations.

Those differences affect both price and what the customer can do with the system.

![Rental and ownership costs meet at 12,500 used GPU hours under the example’s stated cost assumptions.](https://ai.engineer/api/talk-essay-media/5e8a83262be75b7f09a459aab8223f01c29190bdb7354a224c00f7afe5bb6ae5)

Rental and ownership costs meet at 12,500 used GPU hours under the example’s stated cost assumptions.

The value of ownership depends on the workload. Physical access, confidential compute, network topology and the ability to reconfigure a system can matter enough to justify more control and expense. Renting may carry an opportunity cost if the provider's configuration prevents the business from doing something valuable. A comparison based only on GPU-hour prices would miss that cost.

Utilization changes the comparison too. Mustafa describes an illustrative ownership-versus-rental graph and asks whether a $100 million commitment over three years approaches the cost of building and financing a cluster—or even buying the operator. Time and operating costs still belong in that calculation. The graph offers a way to think about the choice, without establishing a universal break-even utilization: sufficiently heavy use can make ownership attractive, while the expense and work of owning the stack remain consequential.

These decisions reach directly into startup runway. Access and payment terms determine how much cash remains for the rest of the company, giving venture investors a practical reason to help portfolio companies find and finance suitable capacity. Procurement can improve the underlying investment by matching the purchase to the company's needs. That leads into the market's next question: who finds the capacity, compares the offers and carries the risks between buyer and supplier?

### What brokers, benchmarks and financing providers contribute

[13:27](https://www.youtube.com/watch?v=GCse4OMzY8U&t=807s)

Several kinds of intermediary work around the GPU hour, solving different parts of a transaction:

- **Price discovery:** Indexes, data and benchmarks provide reference prices. Mustafa describes benchmarks as possible foundations for future futures contracts, subject to their specifications and approval.
- **Physical matching:** Auctions, aggregators, on-demand marketplaces, pooled capacity services, bare-metal resellers and neoclouds connect available compute with buyers. Some pools divide capacity into smaller purchases under their own standardized contracts.
- **Desks and brokers:** Relationships and visibility into bilateral quotes help locate capacity and negotiate execution.
- **Credit and financing:** Equipment lenders, lessors and specialist credit providers finance assets or payment obligations.

![The GPU-hour diagram groups price discovery, physical matching, networks, brokers, financial markets and credit above an insurance and operations layer.](https://ai.engineer/api/talk-essay-media/47e4a01d3a606d49239575b77db2b31e1740e3eb5909070f9b6a6c71c9104616)

The GPU-hour diagram groups price discovery, physical matching, networks, brokers, financial markets and credit above an insurance and operations layer.

A broker's spread can reflect more than access to a contact list. Consider a buyer that needs compute immediately and is relatively insensitive to price. A desk that sees comparable specifications across several quotes may find a cluster that can be ready in time, negotiate the source and charge for filling the urgent need. If it also carries risk or undertakes to operate the cluster for the buyer, the spread must compensate it for those responsibilities.

Financial structures address another set of needs. Mustafa describes prepayment facilities coming online and interest in futures, derivatives and options to hedge exposures already on companies' books. Buyers also want flexibility inside physical contracts: after committing for three to five years, could they upgrade when a new chipset arrives, roll the commitment forward or sell capacity back? These are different contractual rights, each changing who bears the cost of uncertain future demand and hardware changes.

### Why two H100 quotes may describe different products

[15:26](https://www.youtube.com/watch?v=GCse4OMzY8U&t=926s)

Every deal being bespoke creates work before anyone can compare prices. Specification sheets arrive in different formats. A bill of materials can change when a part is missing or out of stock, and equipment manufacturers can have long lead times. Even a familiar word can hide a material difference: does “memory” mean GPU VRAM, CPU memory or additional storage? The answer changes which workloads fit, what the buyer needs and how the offer should be priced.

This is the market's non-fungibility problem: one cluster cannot automatically substitute for another. A chip SKU may carry much of the pricing weight, yet other specifications and delivery conditions can make the quote applicable only to that particular cluster. The standardization question is therefore which attributes can be made comparable and which must remain part of a specific product description. Without that distinction, a common GPU label gives a false impression of interchangeability.

What does a public H100 price comparison actually tell a buyer? Mustafa presents four references: Silicon Data’s daily neocloud index, SemiAnalysis’s spot-contract composite, Compute Desk’s U.S. one-year reserved reference and Ornn’s H100 SXM on-demand transactions. A seller can use these alongside private transaction experience, then adjust for the cluster and the terms. Prices can change daily or faster, bids can be shopped around, and creditworthiness affects the offer. None of the four benchmarks is an executable quote for a particular cluster.

![Four public H100 price references on October 13, 2026 range from $2.85 to $3.32 per GPU hour. Their contract terms and methods differ.](https://ai.engineer/api/talk-essay-media/db12de6da05ea46dda180c9d8e23b8221312346f07106ae08f130ab9a71fed39)

Four public H100 price references on October 13, 2026 range from $2.85 to $3.32 per GPU hour. Their contract terms and methods differ.

### A price hedge cannot deliver a cluster

[17:36](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1056s)

The useful contract starts with the exposure a business needs to manage. An inference cloud selling a fixed-price service while paying a floating GPU bill faces price risk: its input cost can move without a matching change in customer revenue. Locking capacity for years can reduce part of that exposure, but it creates another obligation—the company pays for hours even when it cannot use them.

![The slide lists bridge capacity contracts, buyback and sell-back clauses, offtake and hedging tools. Its final bullet distinguishes a price hedge from securing capacity.](https://ai.engineer/api/talk-essay-media/2588e0b3bd7a3222e695741c2fd6bbebfc879d0f87159cc97e8f964dc57f8d45)

The slide lists bridge capacity contracts, buyback and sell-back clauses, offtake and hedging tools. Its final bullet distinguishes a price hedge from securing capacity.

The available structures solve different problems:

- **Long-term capacity commitments:** Secure capacity and reduce some price exposure, while creating the possibility of paid but idle hours.
- **Resale rights:** Offer a way to recover value from spare capacity, provided another buyer can actually purchase those hours.
- **Financial hedges:** Address movements in a reference price. A cash-settled hedge cannot deliver the required cluster or guarantee that purchased capacity will be utilized.

Bridging capacity, flexible commitments, resale rights and price protection belong in the same conversation because a company may need several of them. Treating one as a complete solution merely moves the unresolved risk elsewhere.

### Standard terms still need someone able to carry the risk

[18:24](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1104s)

Policy also shapes who can obtain compute and under what conditions. The closing policy section cites a January U.S. Commerce Department change moving qualifying H200-class license reviews for China and Macau to case-by-case review with conditions. It also cites the U.S. AI Action Plan's support for open-source and open-weight AI and its recommendation to improve compute financial markets so startups and academics can access large-scale computing power. These are policy examples presented in the recording; eligibility remains conditional, rather than a general promise of access.

![The closing slide identifies middle-market buyers, balance sheets that can bridge buyer and supplier needs, and standard specifications and contract terms.](https://ai.engineer/api/talk-essay-media/f125a6491a43dfaeb5cd4f44898f70d00ad0ee2d1a51a86e441e4b950a42b4c8)

The closing slide identifies middle-market buyers, balance sheets that can bridge buyer and supplier needs, and standard specifications and contract terms.

The commercial opportunity returns to the middle market. Growing companies have real workloads, but their financing, timing and specification needs often do not fit the contracts that hyperscalers and frontier labs can sign. Suppliers want long commitments to support a build-out. Buyers want to expand in smaller blocks as volatile demand becomes clearer, avoiding payment for capacity they cannot yet use.

Common capacity contracts, specifications, terms and definitions would make offers easier to compare. They would not eliminate the mismatch in commitments. The closing proposal also requires participants with strong credit who can assess the risks and carry the commitments on strong balance sheets uncorrelated with the compute market.

That role requires understanding delivery and workloads as well as quoting a GPU hour. Hardware expertise matters because the intermediary must judge whether the promised system will serve the buyer's needs; financial strength matters because buyer demand and supplier obligations do not line up neatly. The closing invitation—“let's build this together”—is a proposal to combine those capabilities into a market that growing companies can actually use.

## Key takeaways

- Long-term GPU capacity purchases combine hardware, delivery and commercial terms. Public GPU-hour prices provide references, while substantial commitments still trade through private negotiation.
- Lower token costs can encourage enough additional usage to sustain hardware demand. The talk offers this as a possible explanation for falling token prices alongside firm compute prices.
- Payment terms can determine startup runway: in the illustrative case, a $40 million raise leaves only $10 million after a 30% prepayment on a $100 million compute contract.
- Physical commitments, resale rights and financial hedges address different risks. A cash-settled price hedge cannot supply a cluster or make idle capacity useful.
- A more useful middle market needs both comparable contracts and participants able to finance and assess the gap between suppliers' long commitments and buyers' incremental growth.

## Related talks

- [A Taxonomy for Next-Generation Reasoning Models](https://ai.engineer/talks/jQcsVk0KWiQ-taxonomy-next-generation-reasoning-models) — Explores inference-time scaling and token costs from the model side, adding context for why cheaper individual tasks can coexist with growing demand for compute.

## Slides

### The Dark Pools of Compute

Affaan Mustafa and Alejandro Soumah of Jump Trading introduce compute trading: what it is, how its market is structured, and where they see it heading, including policy.

[0:12](https://www.youtube.com/watch?v=GCse4OMzY8U&t=12s) [Opening slide titled The Dark Pools of Compute, naming Affaan Mustafa and Alejandro Soumah of Compute Trading at Jump Trading.](https://assets.ai.engineer/sites/aiecode2025/sha256/01dfa9c07f8905441fe125700678f87a1df7248f200a87e0118b3cc5978eaee1.webp)

OPENING
01

The Dark Pools of Compute

Affaan Mustafa · Alejandro Soumah | Compute Trading at Jump Trading

Interface text:
CC

### What do I mean by dark pool?

Here, dark pool describes a fragmented compute market: private quotes, brokers and long bilateral contracts keep most deals out of public view. The analogy differs from equities’ regulated, structured venues.

[0:51](https://www.youtube.com/watch?v=GCse4OMzY8U&t=51s) [Three bullets contrast regulated equities venues with compute's private quotes, brokers and long bilateral contracts, noting that most of the market happens out of view in fragments.](https://assets.ai.engineer/sites/aiecode2025/sha256/c735c8c50be6bebe0bc9b07c477ec7d4d396824e38d1ea31e9cc7b590b4231cf.webp)

What do I mean by dark pool?

OPENING
02

• Equities: a regulated, structured venue
• Compute: private quotes, brokers and long bilateral contracts
• Most of the market happens out of view, in fragments

Interface text:
cc

### Tokens get cheaper while compute demand grows

LLM prices fall much faster than PC prices did in their investment cycle. Cheaper tokens let teams do more work on the same budget, but the underlying capacity still has to be built and paid for.

[2:38](https://www.youtube.com/watch?v=GCse4OMzY8U&t=158s) [Line chart compares declining personal-computer and LLM price indices over investment cycles. Two bullets say capability costs have fallen sharply while the total bill depends on model, usage and purchased capacity.](https://assets.ai.engineer/sites/aiecode2025/sha256/ea0aa594afe4d1660f4dde6df1751394ea3d6622d0e3ef0bf402939d21a782f7.webp)

Tokens get cheaper while compute demand grows

01 MACRO MARKET STRUCTURE & PLAYERS
03

• The cost of a given capability has fallen sharply.
• The total bill still depends on the model, usage and the capacity being purchased.

Costs of New Technology for End-Use Consumers,
by investment cycle

Legend:
PCE Personal Computer Price Index (1980 ICT Investment Cycle)
LLM Model Price Index* (2022 AI Investment Cycle)
LLM Model Price-to-Intelligence Ratio** (2022 AI Investment Cycle)

Left Y-axis ticks: 0, 20, 40, 60, 80, 100, 120
Right Y-axis ticks: 0, 20, 40, 60, 80, 100, 120
X-axis: Months Since Investment Cycle Start
X-axis ticks: t+0, t+20, t+40, t+60, t+80, t+100, t+120, t+140, t+160, t+180, t+200

Sources: a16z State of Markets II (Sep 2026), slide 34; Goldman Sachs / Commerce (Jul 2026); Demirer et al., JEP (2026).

Interface text:
CC

[2:55](https://www.youtube.com/watch?v=GCse4OMzY8U&t=175s) [Line chart comparing personal-computer prices, LLM model prices and model price-to-intelligence ratios over investment cycles. Bullets distinguish falling capability costs from total spending on models, usage and capacity.](https://assets.ai.engineer/sites/aiecode2025/sha256/4f0569ceaad6389b324b4cb7467f5cd8190822ede63e6f268f2f818449548e48.webp)

Tokens get cheaper while compute demand grows

01 MACRO MARKET STRUCTURE & PLAYERS
03

• The cost of a given capability has fallen sharply.
• The total bill still depends on the model, usage and the capacity being purchased.

Costs of New Technology for End-Use Consumers,
by investment cycle

Legend:
PCE Personal Computer Price Index (1980 ICT Investment Cycle)
LLM Model Price Index* (2022 AI Investment Cycle)
LLM Model Price-to-Intelligence Ratio** (2022 AI Investment Cycle)

Left Y-axis ticks: 0, 20, 40, 60, 80, 100, 120
Right Y-axis ticks: 0, 20, 40, 60, 80, 100, 120
X-axis: Months Since Investment Cycle Start
X-axis ticks: t+0, t+20, t+40, t+60, t+80, t+100, t+120, t+140, t+160, t+180, t+200

Sources: a16z State of Markets II (Sep 2026), slide 34; Goldman Sachs / Commerce (Jul 2026); Demirer et al., JEP (2026).

Interface text:
CC
AI Engineer

### Jevons and the decoupling of token and capacity prices

From July through August 2026, the token-price index falls while H100 rental prices hold up. Cheaper tasks may spur usage that offsets efficiency gains; capacity prices also depend on workload mix and usable-cluster supply.

[4:01](https://www.youtube.com/watch?v=GCse4OMzY8U&t=241s) [Dual-axis chart of H100 GPU rental prices and a token-price expenditure index, with highlighted periods of joint decline and subsequent divergence. Text explains that capacity prices also depend on efficiency, workload mix and usable-cluster supply.](https://assets.ai.engineer/sites/aiecode2025/sha256/d6c222808e226e593733540e13ce15a5d90a48e30d62392db3f97be0b59b1e65.webp)

Jevons and the decoupling of token and capacity prices

01 MACRO MARKET STRUCTURE & PLAYERS
04

Legend:
Ornn H100 GPU Rental Price
Silicon Data Token Price Expenditure Index

Left Y-axis: H100 Rental Price ($/hr)
Left Y-axis ticks: 1.00, 1.50, 2.00, 2.50, 3.00, 3.50
Right Y-axis: Token Price Expenditure Index
Right Y-axis ticks: 0.90, 1.10, 1.30, 1.50, 1.70, 1.90, 2.10, 2.30, 2.50
Timeline: Dec ’25, Jan ’26, Feb ’26, Mar ’26, Apr ’26, May ’26, Jun ’26, Jul ’26, Aug ’26, Sep ’26

Chart annotations:
Token and compute prices decline together
Token prices decline further but compute prices remain robust

Lower cost per task can expand usage, while capacity prices also depend on efficiency, workload mix and the supply of usable clusters.

Source footnote:
a16z State of Markets II, slide 34; Ornn, Silicon Data, Bloomberg, compiled by Citadel Securities; data through August 2026.

### Four ingredients

Usable GPU compute requires chips, power, buildings and grid connections, and capital to come together. The IEA projects electricity use across all global data centres rising from 485 TWh in 2025 to 950 TWh in 2030.

[5:06](https://www.youtube.com/watch?v=GCse4OMzY8U&t=306s) [Four ingredients for compute infrastructure: chips, power, buildings and grid connections, and money. The power bullet cites an IEA projection of global data-centre electricity use increasing from 485 TWh in 2025 to 950 TWh in 2030.](https://assets.ai.engineer/sites/aiecode2025/sha256/f92820ad2fe9cbb40df8af99ad5ad0a7d75eba59a61ab426499dcd8f11a32495.webp)

Four ingredients

01 MACRO MARKET STRUCTURE & PLAYERS
05

• Chips
• Power: the IEA projects global data-centre electricity use rising from 485 TWh in 2025 to 950 TWh in 2030
• Buildings and grid connections
• Money

Source: IEA, Key Questions on Energy and AI (April 2026); the projection covers all global data centres.

Interface text:
CC

### How the compute market fits together

Compute moves from chip and system vendors through sites and operators to capacity providers and users. Firms can span layers; lenders and investors fund buildout, while brokers and marketplaces connect capacity offers with buyers.

[6:14](https://www.youtube.com/watch?v=GCse4OMzY8U&t=374s) [Four linked boxes show chip and system vendors, sites and operators, capacity providers, and compute users. Supporting text describes lenders and investors funding buildout and brokers and marketplaces connecting capacity offers with buyers.](https://assets.ai.engineer/sites/aiecode2025/sha256/3b3bcd8259d3aabb235998fd57abfaea29d81915f7d36309094efccecb94640f.webp)

How the compute market fits together

01 MACRO MARKET STRUCTURE & PLAYERS
06

Chip and system vendors
GPUs, memory, servers and networking

→

Sites and operators
Power, space, assembly and management

→

Capacity providers
Hyperscalers, neoclouds and dedicated clusters

→

Compute users
Labs, inference clouds and enterprises

Lenders and investors support the hardware and site buildout.

Brokers and marketplaces help connect specific capacity offers with buyers.

Schematic: firms can span several layers. Public context: CoreWeave 2025 10-K; IEA (April 2026); SemiAnalysis AI Cloud TCO Model.

Interface text:
CC

### The financing chain

In this illustration, a 36-month hardware loan outlasts a 24-month customer contract, leaving 12 months of re-let risk. Lenders rely on contracted revenue; credit support and payment timing shape what gets financed.

[8:12](https://www.youtube.com/watch?v=GCse4OMzY8U&t=492s) [Illustrative timeline comparing 36-month hardware financing with a 24-month customer contract, leaving 12 months of re-let risk. Text explains contracted revenue, term mismatch, credit support and payment timing.](https://assets.ai.engineer/sites/aiecode2025/sha256/f1066e8bbc076ca3e4ffa2159f78f84e7cdd771add83f70d86c4acbb22dfe613.webp)

The financing chain

01 MACRO MARKET STRUCTURE & PLAYERS
07

Lenders finance the buildout against contracted revenue.

The customer’s term can end before the financing does.

Credit support and payment timing affect what gets financed.

A three-year loan with two-year offtake

Hardware financing
Customer contract
Re-let risk

Timeline: 0m, 12m, 24m, 36m

Illustrative term mismatch, not an observed transaction. CoreWeave filings; Liquid Compute; Mittal / Mustafa, Handmade.

Interface text:
CC

### The startup math (illustrative)

In this illustration, a $30M prepayment on a $100M, three-year contract leaves $10M of a $40M raise to operate. Capacity arrives in blocks: spare capacity may serve short-term buyers; on-demand can bridge later shortfalls.

[9:23](https://www.youtube.com/watch?v=GCse4OMzY8U&t=563s) [Illustrative example: raising $40 million and prepaying 30 percent of a $100 million, three-year contract leaves $10 million to operate. A chart contrasts rising workload demand with stepped capacity commitments, showing unused capacity and an eventual shortfall bridged on demand.](https://assets.ai.engineer/sites/aiecode2025/sha256/ae217b4d8e3bec4fde38a3f567dc002b338208203a3b7cb20027a1dc384ce10c.webp)

The startup math (illustrative)

01 MACRO MARKET STRUCTURE & PLAYERS
08

Raise $40M

Need a $100M contract over 3 years

30% prepayment = $30M

$10M left to run the business

An inference cloud commits capacity in blocks

Legend: Workload demand; Committed capacity

Unused capacity can serve short-term buyers

On-demand bridges the shortfall

X-axis: Time

Shading legend: Capacity above demand; Demand above capacity

Illustrative, not a forecast. Resale requires buyers, matching specifications and contractual permission.

Interface text:
CC

### Rent versus own

Full-stack operators arrange hardware, site and operations. Colocation splits those duties; offtake buyers hold capacity rights while providers supply hardware, site and operations. Responsibilities vary by agreement.

[10:33](https://www.youtube.com/watch?v=GCse4OMzY8U&t=633s) [Table comparing full-stack build and operation, hardware ownership in managed colocation, hardware ownership in self-managed colocation, and capacity offtake contracts by what is owned and who supplies the site and operations.](https://assets.ai.engineer/sites/aiecode2025/sha256/5601b7d3532e550618239d141f98f75aa53423da60a207e81fb9e716124c9012.webp)

Rent versus own

01 MACRO MARKET STRUCTURE & PLAYERS
09

Operating model | What you own | Who supplies the site and operations

Build and operate the full stack | Hardware and site | Your team arranges power, permits, assembly and management

Own hardware in managed colo | Servers | Colo supplies power and space; a partner manages hardware

Own hardware in self-managed colo | Servers or components | Colo supplies power and space; assembly is outsourced and your team operates

Capacity offtake contract | Capacity rights | Provider supplies the hardware, site and operations

Schematic: responsibility, access and service boundaries vary by agreement. Source framework: SemiAnalysis AI Cloud TCO Model.

Interface text:
cc

### How ownership and rental compare as utilization rises

In this hypothetical, undiscounted example, ownership costs $25,000 plus $1 per used GPU-hour versus $3 for rental. Costs meet at 12,500 used hours; ownership is cheaper beyond that. Term contracts may charge for idle hours.

[11:43](https://www.youtube.com/watch?v=GCse4OMzY8U&t=703s) [Hypothetical total-cost chart comparing rental at $3 per used GPU-hour with ownership at $25,000 net fixed cost plus $1 per used hour. Lines intersect at 12,500 used hours; notes explain that assumptions determine the crossover and term contracts may charge for idle allocated hours.](https://assets.ai.engineer/sites/aiecode2025/sha256/f88a1fae28887c0c91bf5cc7003995e347f5e063a84734f0768a47507da24eb0.webp)

How ownership and rental compare as utilization rises

01 MACRO MARKET STRUCTURE & PLAYERS
10

Total cost per GPU over the horizon
Y-axis ticks: $0k, $25k, $50k, $75k
X-axis: Used GPU-hours
X-axis ticks: 0k, 12k, 24k
Legend: Rental; Ownership
Intersection label: 12,500 used hours

With $25,000 of net fixed ownership costs, $1/hour of operating costs and $3/hour rental, the curves meet at 12,500 used hours.

The assumptions determine where they cross.

Hypothetical, undiscounted same-workload example. It compares pay-as-used rental; a term contract can charge for idle allocated hours.

Interface text:
cc

### A new dynamic

Compute procurement shapes a growing company’s runway. Investors therefore care about payment terms and access to capacity—and can help portfolio companies secure and fund suitable compute.

[12:40](https://www.youtube.com/watch?v=GCse4OMzY8U&t=760s) [Two bullets explain that compute procurement affects a growing company’s runway and that investors care about procurement terms and access to capacity.](https://assets.ai.engineer/sites/aiecode2025/sha256/48717710dfced4084d4419915a06acffa47990401d7a96be5506336c920994a1.webp)

A new dynamic

01 MACRO MARKET STRUCTURE & PLAYERS
11

• Compute procurement affects how much runway a growing company has
• Investors have an interest in the terms as well as the company’s access to capacity

Interface text:
cc

### The players around the GPU-hour

The GPU-hour market separates matching physical capacity from trading price risk. Brokers arrange deals, lenders fund hardware, and an insurance-and-operations layer spans these roles. Roles can overlap.

[13:28](https://www.youtube.com/watch?v=GCse4OMzY8U&t=808s) [Boxes organize GPU-hour market roles into price discovery, physical matching, decentralized networks, desks and brokers, financial markets, credit and financing, and insurance and operations.](https://assets.ai.engineer/sites/aiecode2025/sha256/516e1acd7e664542d0eda242a69a7221ae9f3053f4184d77888572888e70c296.webp)

The players around the GPU-hour

02 MICROSTRUCTURE & PLAYERS
12

Price discovery
Indices and data
Benchmarks, transaction data and comparable price references

Physical matching
Venues, auctions and aggregators
Matching capacity offers with suitable workloads

Decentralized networks
Distributed supply and settlement
Coordinating distributed resources and access to available capacity

Desks and brokers
Bilateral quotes and execution
Sourcing, negotiation and carrying or placing risk

Financial markets
Price risk and reference contracts
Exchanges and derivatives
Separate from physical supply

Credit and financing
Funding hardware and commitments
Equipment lenders and lessors
Specialist credit providers

Insurance and operations
Protection, telemetry, metering, billing and collateral services

Functional view adapted from a public September 2026 compute-market map. Roles can overlap; not exhaustive.

Interface text:
CC

### Every deal is bespoke

GPU offers need back-and-forth before buyers can compare them: spec formats and bills of materials differ, and even memory can mean VRAM, CPU memory or storage.

[15:48](https://www.youtube.com/watch?v=GCse4OMzY8U&t=948s) [Four bullets identify different specification formats, mismatched bills of materials, terms with different meanings, and extensive back-and-forth before offers can be compared.](https://assets.ai.engineer/sites/aiecode2025/sha256/bc3069cf2d64bb84c8f8b42dad1a4cccac8ffd744d20b6150d5bac99efe381a3.webp)

Every deal is bespoke

02 MICROSTRUCTURE & PLAYERS
13

• Different spec sheets and formats
• Bills of materials that don’t match
• The same term means different things
• Lots of back-and-forth before anyone can compare

Interface text:
CC

### Pricing is a negotiation

H100 references on 13 October 2026 span $2.85–$3.32 per GPU-hour, but use different methods. Actual offers shift with daily repricing, bid shopping and buyer creditworthiness; published benchmarks are not quotes.

[16:46](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1006s) [Four bullets describe changing prices, starting quotes, bid shopping, and creditworthiness. A comparison lists four public H100 price references dated October 13, 2026, noting differing methods and that benchmarks are not quotes.](https://assets.ai.engineer/sites/aiecode2025/sha256/93045454dd04ee16318e88db896f6554d2e57a3142a3289e9ad7b55970f57de4.webp)

Pricing is a negotiation

02 MICROSTRUCTURE & PLAYERS
14

• Prices can update daily or faster
• A quote is a starting point
• Bids get shopped to push the price
• Creditworthiness changes the offer

Public H100 references, 13 October 2026

Silicon Data    $2.85
Standardized neocloud daily index

SemiAnalysis    $3.32
H100 spot-contract composite

Compute Desk    $2.96
US one-year reserved reference

Ornn    $3.17
H100 SXM on-demand transactions

USD per GPU-hour; the definitions and methods differ.

Sources: Silicon Data; SemiAnalysis; Compute Desk / General Index; Ornn. Published benchmarks are not quotes.

Interface text:
CC

### Pricing is a negotiation

H100 benchmarks on 13 October 2026 span $2.85–$3.32 per GPU-hour, but use different methods. Quotes start a negotiation: bids get shopped, prices move daily or faster, and creditworthiness changes the offer.

[16:55](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1015s) [Slide describing changing prices, negotiable quotes, bid shopping and creditworthiness, alongside four dated public H100 pricing references with differing methodologies.](https://assets.ai.engineer/sites/aiecode2025/sha256/763d19b7bc468efb932241d7419c139f5381b2ed326b825124f3456c45b73cc8.webp)

Pricing is a negotiation

02 MICROSTRUCTURE & PLAYERS
14

• Prices can update daily or faster
• A quote is a starting point
• Bids get shopped to push the price
• Creditworthiness changes the offer

Public H100 references, 13 October 2026

Silicon Data    $2.85
Standardized neocloud daily index

SemiAnalysis    $3.32
H100 spot-contract composite

Compute Desk    $2.96
US one-year reserved reference

Ornn    $3.17
H100 SXM on-demand transactions

USD per GPU-hour; the definitions and methods differ.

Sources: Silicon Data; SemiAnalysis; Compute Desk / General Index; Ornn. Published benchmarks are not quotes.

Interface text:
AI Engineer

### Why new structures

An inference cloud can sell at a fixed price while its GPU bill floats. Long commitments reduce price exposure but risk idle hours; resale needs buyers, and a financial hedge cannot deliver a cluster or guarantee utilization.

[17:39](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1059s) [Slide covering funding costs, flexible and bridge capacity contracts, resale clauses and hedging, noting that a price hedge does not secure capacity.](https://assets.ai.engineer/sites/aiecode2025/sha256/fa4c543edc95cf87e21e136ed14b9baa120343edb694376522a056b657511aed.webp)

Why new structures

02 MICROSTRUCTURE & PLAYERS
15

• Funding costs can change the comparison between term and on-demand capacity
• Flexible and bridge capacity contracts
• Buyback and sell-back clauses, offtake, hedging tools
• Short-term resale needs filled hours; a price hedge does not secure capacity

Additional context: Liquid Compute, The Tenor Trade and Hedging Compute Prices (September 2026).

### Policy

Export rules shape who can access compute; the US AI Action Plan links startup and academic access to compute finance. Power and permitting affect when that capacity becomes usable.

[18:22](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1102s) [Policy slide discussing conditional export reviews, open-weight AI and compute access, and the effects of power and permitting on usable capacity.](https://assets.ai.engineer/sites/aiecode2025/sha256/69e9372cc427d9d8f194bf489d5cadf97a72f91ff031bdc0085b9a0f6f061d8f.webp)

Policy

03 POLICY · ALEJANDRO
16

• Export controls: qualifying H200-class exports to China and Macau moved to conditional case-by-case review in January 2026
• Open-weight AI: the US AI Action Plan explicitly links access for startups and academics to improving the financial market for compute
• Power and permitting affect when capacity becomes usable

Sources: BIS, 91 FR 1684 (Jan 2026); America’s AI Action Plan, pp4–5 (Jul 2025); IEA (Apr 2026).

### The opportunity

Middle-market buyers want capacity in smaller blocks; suppliers want long-term commitments to fund build-outs. Strong balance sheets and standard capacity contracts could bridge that mismatch.

[19:15](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1155s) [Closing slide identifying the middle market, balance sheets bridging buyer and supplier needs, and standardized capacity contracts, above an AI Engineer branded footer.](https://assets.ai.engineer/sites/aiecode2025/sha256/8858e98de000892a656173b6de7ec303fb9a286804438605a4295adb27d49d40.webp)

The opportunity

04 OPPORTUNITY & CLOSING · ALEJANDRO
17

• The middle market
• Balance sheets that can bridge buyer and supplier needs
• A standard capacity contract: common specs, terms and definitions
• Let’s build it together

AI Engineer

Interface text:
CC

### The opportunity

Middle-market buyers need capacity in smaller blocks; suppliers want long-term commitments. Strong balance sheets can bridge that mismatch, while standard capacity contracts give both sides common specs, terms and definitions.

[19:19](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1159s) [Closing slide calling for middle-market financing, balance sheets that bridge buyer and supplier needs, common capacity-contract specifications and terms, and collaboration.](https://assets.ai.engineer/sites/aiecode2025/sha256/275e6b2a56efed29d12bcadffb42a7ef4bc5f87df3092102603d1a846dbb6d1e.webp)

The opportunity

04 OPPORTUNITY & CLOSING · ALEJANDRO
17

• The middle market
• Balance sheets that can bridge buyer and supplier needs
• A standard capacity contract: common specs, terms and definitions
• Let’s build it together

Interface text:
CC

## Full transcript

[0:12](https://www.youtube.com/watch?v=GCse4OMzY8U&t=12s) Um, hey everyone. My name is Affaan, and, um, I'm a lead for compute trading at Jump Trading, which is a proprietary trading firm, alongside my co-lead over here, Alejandro. So today we're gonna be looking over some of the market structure as it applies to compute trading, what compute trading even is, um, what we can see it becoming. And I think we'll start with some macro, look at microstructure, and then talk a little bit about where we see things in policy.

[0:44](https://www.youtube.com/watch?v=GCse4OMzY8U&t=44s) So the title of the talk is The Dark Pools of Compute, but some of you guys are in finance, so you might be confused as to why I'm using this terminology. Uh, what do I mean by dark pool here? Essentially, what I'm referring to is an unlit market. Um, right now most compute trades in blocks, physical compute here, and these are private quotes. We have a lot of brokers, and these are long bilateral contracts. Now, at the end of the d-day, these are really commercial agreements, so we're using the term trading here to kind of like look at how it's transacted between

[1:14](https://www.youtube.com/watch?v=GCse4OMzY8U&t=74s) parties. These commercial agreements tend to be some mixture of an MSA, an SLA, and an order form, cluster-specific, uh, with hardware-specific, uh, language in each kind of, um, in each order. And but when we're looking at pricing and looking at some of these other kind of, uh, metrics, a lot of the time most of the talk is around the SKU. So what the chip itself is, the baseboard. So H100, H200, B300, uh, et cetera. And so, um,

[1:44](https://www.youtube.com/watch?v=GCse4OMzY8U&t=104s) a lot of the negotiation here happens kind of outside the public view. There's no, you know, like exchange for physically traded compute that's widely, you know, adopted and, you know, used for everyone goes there to like take a look and trade, uh, compute. These are done in chats, right? They're out of the public view. Um, there's public marketplaces, there's price references, there's a lot of, um, you know, on-demand marketplaces. However, for these actual tenored, so, you know, on a given term for a fixed amount of capacity, uh, with a

[2:14](https://www.youtube.com/watch?v=GCse4OMzY8U&t=134s) certain ter-- total contract value, these trade over Slack messages, X messages, Telegram group chats. There's a bunch of OTC kind of channels that are all over in these kind of, I wanna call them like Neo-Neochat. Um, and some of you might be familiar with commodities and how a lot of, you know, there's chatting happening on ICE Chat, Bloomberg chat, et cetera. This is kind of where, where the, where the kind of ma-markets here lie. And so I want to look a little bit here on the macro market structure and a little bit on the

[2:44](https://www.youtube.com/watch?v=GCse4OMzY8U&t=164s) players here as well. And so this was a recent kind of, uh, a16z. Some of you guys might have seen this, so I'm, I'm putting this out there because, you know, you can go on a16z's, uh, actual X and take a look at this graph yourself. What we're looking at is that the cost of a given capability has fallen sharply. Um, the bill of course still depends on the model usage, capacity, and other factors such as speed is a big thing right now with the new co-- uh, ChatGPT models. Um, cheaper tokens unlock more usage and a

[3:14](https://www.youtube.com/watch?v=GCse4OMzY8U&t=194s) team can do more work for the same budget. So what we're seeing is that a lot of teams are simply just using way more tokens. So we have this latent demand that hasn't really been hit yet, and people are just going to use more with the same budget that they were using before when models cost more, um, and they're just spending more tokens. Um, so applications that were too expensive become possible. Um, however, as we're kind of talking about compute here, the capacity underneath still has to be built and paid for, um, even as the cost of a particular task comes down. And so the

[3:44](https://www.youtube.com/watch?v=GCse4OMzY8U&t=224s) important distinction is between the price of the intelligence, the number of tasks people want to run, and the amount of hardware needed to serve them.

[3:53](https://www.youtube.com/watch?v=GCse4OMzY8U&t=233s) And so this is where we get to kind of the Jevons paradox and a little-- some of the stuff that we've been looking at lately, where we had a very interesting thing where, you know, intuitively you'd assume, um, that compute and inference kind of like rises together, correlates very, you know, very closely as you can see. You can see there's a little bit of a decoupling that happened and token prices declined further, but compute prices re-remain robust. And so like what, what can be an explanation for this? Now, there's a lot of like potential explanations. One

[4:23](https://www.youtube.com/watch?v=GCse4OMzY8U&t=263s) here is the whole Jevons paradox is that even though token prices and inference is getting extremely efficient to serve on these new, uh, baseboards and, and, and GPUs, people are just consuming a lot more. And so the demand for compute rises and people just consume a lot more tokens even if the cost per, uh, cost per million, uh, uh, for input and for caching and, you know, for output becomes a lot more efficient. And so making the tasks cheaper, encouraging the additional use to outweigh resource saving per task. Uh, for inference clouds, that's going

[4:53](https://www.youtube.com/watch?v=GCse4OMzY8U&t=293s) to mean more applications, serving more tokens, uh, bigger market share. Um, and for app layer individuals, that means you guys can serve to a lot more kinda customers. So anyone who's building utilizing the token, uh, infrastructure.

[5:08](https://www.youtube.com/watch?v=GCse4OMzY8U&t=308s) And so what really goes into, um, you know, the macro structure. Now like this is very, you know, bare bones here, but you've got your chips, which we've been talking about. Uh, you have power, you have buildings, grid connections, and then of course you need a lot of capital to get these build-outs going, and then that's kind of been the whole talk here. And so, um, chips, powers, buildings, money, they all have to come together before, you know, you can even get to an hour of like actual usable GPU-per-hour compute. And so the IEA April update projects that the glo-global data

[5:38](https://www.youtube.com/watch?v=GCse4OMzY8U&t=338s) center electricity is gonna rise from four hundred and eighty-five terawatt-hours in '25, um, to nine hundred and fifty in 2030. And so a lot of the talk right now is around, um, build-outs, about getting the power grid set up. Will we have enough power grid here? Is there other energy sources that we can tap into? Um, where are we, you know, able to get the land to actually be able to do this? And so colocation sites and data center operators and people who are doing the financing, the build-outs or the lenders themselves are all kind of looking into this and that type of, uh, everything

[6:08](https://www.youtube.com/watch?v=GCse4OMzY8U&t=368s) that's upstream essentially of the compute is going to then flow down into the compute, uh, costs and also usage. And so what does this like actual macro map look like underneath it? Quote, "Chip and system vendors supply the GPU servers and networking. Sites and operators supply the power, um, space assembly and management." Right here, GPUs, memory servers, and networking. Um, capacity providers, these are hyperscalers and neoclouds and any other operator that's like operating similar to how a neocloud would

[6:37](https://www.youtube.com/watch?v=GCse4OMzY8U&t=397s) operate, reselling bare metal capacity. And then the users. So who's the actual end user? The startups, the inference clouds, uh, labs, neo labs, and also anyone on the app layer, uh, and enterprises for anyone doing, um, open-weight stuff as well. And, um, these boxes, like I've kind of boxed them out here, but, you know, there's... People can be all of the above. A chip and system vendor, like someone could do a full end-to-end build-out and also be the end user of everything. And so there's a

[7:07](https://www.youtube.com/watch?v=GCse4OMzY8U&t=427s) little bit of, um, variation that, that, that occurs here.

[7:18](https://www.youtube.com/watch?v=GCse4OMzY8U&t=438s) And so as I was kind of mentioning about the inference clouds before, they might rent a cluster, and they'll serve tokens off of it, and they'll use that revenue to pay off the cluster. Um, and so that might be a use case for the actual inference cloud themselves. However, for hyperscalers, frontier labs, and larger neoclouds, they can make substantial commitments because they have the ability to get into the large equity, they have the ability to take on more debt, and, uh, they have the ability to do long horizon kind of, um, deals. So ten-year deal, OpenAI

[7:47](https://www.youtube.com/watch?v=GCse4OMzY8U&t=467s) contracted out, um, et cetera. So of that nature. However, anyone in the middle market isn't actually able to do this. Who's underwriting a Series B startup with enough ca-capacity to get into two hundred million plus contract over three years for compute? How many equity rounds will they have to raise, right? Will they have to use their entire equity round to fund just the actual cluster, and then they don't have any remaining capital left to use for actual costs?

[8:14](https://www.youtube.com/watch?v=GCse4OMzY8U&t=494s) And so we're looking into a little bit of the financing chain here and how that plays in. So lenders are financing the build-out, um, against contracted revenue even in many cases for inference clouds. Um, and so they want to finance the build-out, and they want confidence that their loan will get repaid. Wha-- I mean, like, how are you gonna underwrite a startup, like a private company that doesn't have, like, much else going for them right now? Like, what are the... You know, they're not publicly traded. There's no that much cash flows, balance sheet income statement. They might even be pre-re-- pre-profit, pre-revenue, et cetera. Um, and so what they're looking at is like, can the person get contracted revenue over a certain scale?

[8:45](https://www.youtube.com/watch?v=GCse4OMzY8U&t=525s) And so CoreWeave's public filings give large-scale example of infrastructure financing supported by take-or-pay customer contracts. And so this diagram kind of shows the smaller, um, illustrative example here. Um, and the mismatch is that hardware financing, they want very long terms, but customers are only able to afford or don't really want to, like, get into a long-term financing over thirty-six months. They have to put up on up to fifteen to twenty-five percent is what CoreWeave, um, has quoted that they've seen their median, um, prepayment to be fifteen to twenty-five percent upfront,

[9:15](https://www.youtube.com/watch?v=GCse4OMzY8U&t=555s) um, on a hundred million PCV cluster. That's fifteen to twenty-five million dollars that a, um, startup has to put upfront. That might be their entire round.

[9:25](https://www.youtube.com/watch?v=GCse4OMzY8U&t=565s) And so this is kind of a look at what I'm, you know, referencing here. It's like you're raising forty million. You need a hundred million contract over three years. You put down thirty percent prepayment as an illustrative example. You've got ten per-- ten million left to run the business. Um, and so you might have this kind of step function as you can see here. So that's your contracted capacity and what you can actually... And the demand is like, you know, uh, below that. And so now every time that you have to get into a new cluster, it's like tied to your equity round or whatever else it may be, but you might have to cap out and turn away customers. So what are your options?

[9:56](https://www.youtube.com/watch?v=GCse4OMzY8U&t=596s) Do you either turn away customers until you can get into your next, um, you know, equity round so you can put down another thirty percent prepayment so you can get next little bit of capacity going? Or do you use on-demand to close the area under the gap and bridge the shortfall? That's like an, that's like an option now. Now you have to look into the economics. What does that look like? What's the difference between, you know, maybe instead what they do is they go to a lender to finance the prepayment component. But if the yield on the prepayment component that they're getting financed is less than, um, them just buying it out on

[10:25](https://www.youtube.com/watch?v=GCse4OMzY8U&t=625s) demand at that time, if they're able to get into that amount of capacity, then does the economics make sense there?

[10:36](https://www.youtube.com/watch?v=GCse4OMzY8U&t=636s) And so here are some of the operating models that we've seen. It's like you can build and operate the full stack. Hardware, site, your team is gonna arrange power permit, the assembly, the management. You can also own servers in a colocation facility with more or less the assembly and operation handled by a partner. Um, on the other end, an offtake contract buys defined capacity, and the provider is going to own and operate the asset, and they might not give you complete root access to the bare metal, or there might be kind of like levels to how much access you get depending on if it's tenant, multi-tenant,

[11:06](https://www.youtube.com/watch?v=GCse4OMzY8U&t=666s) and et cetera type of like configurations that can also affect then the pricing. Um, and so some businesses are gonna value the physical access. Some bus-businesses are gonna value, um, you know, confidential compute. Uh, you know, they might value certain topology changes or the ability to reconfigure the system. And so if that is something for a business, then they might actually want to have more and more control and build and operate the stack, and the economics for them might make sense given the opportunity cost they would have to give up, um, if they were to just take rental offtake contracts. So they might be willing to

[11:36](https://www.youtube.com/watch?v=GCse4OMzY8U&t=696s) pay more there.

[11:44](https://www.youtube.com/watch?v=GCse4OMzY8U&t=704s) And so here's how, like, ownership and rental... This is a very, like, crude graph. Um, how ownership and rental can compare as utilization r-rises. Um, and so what we're looking at is that usually, you know, an offtake contract or the-- these offtake contracts are getting to the cost of actually almost building out and financing your own cluster. And it's not... 'Cause I mean, if you're spending a hundred million over a three-year period, then maybe if the economics make sense, it would've made more sense to simply buy out the compute operator itself outside of the time cost and then obviously the operational cost that

[12:14](https://www.youtube.com/watch?v=GCse4OMzY8U&t=734s) might exist. So there's a, there's a equilibrium that exists, um, that, that has to be followed where You know, the ownership costs outweigh, um, the, um, you know, rental costs, and the rental costs should sit below the ownership costs. But if you're utilizing it at such a rate and, uh, you know, in, in such variable kind of increments, it might be best to just own the stack outright.

[12:42](https://www.youtube.com/watch?v=GCse4OMzY8U&t=762s) And so we're seeing this new dynamic that compute, uh, the compute procurement affects how much runway a growing company actually has. And so this also gives, uh, investors an interest in compute procurement. So we're kind of seeing this. Um, access and payment terms affect how much runway their portcos are going to have from a VC perspective. Uh, so helping a company buy a suitable capacity can improve the underlying investment. And they're very connected, right? So they're able to procure, and they're able to actually find, uh, compute capacity for their portcos and, and fund that. And so

[13:12](https://www.youtube.com/watch?v=GCse4OMzY8U&t=792s) that also brings another interesting, uh, aspect to the macro structure, um, into the microstructure is that who helps place the capacity and, uh, how does this offer get compared and what risks these contracts can, can transfer.

[13:27](https://www.youtube.com/watch?v=GCse4OMzY8U&t=807s) So we have a lot of different players around this GPU hour that are helping kind of shape this, this market here. You've got your price discovery individuals, so people with, uh, indexes, data, the benchmarks that we're all looking at right now and referencing, um, which would be used to, you know, in the future, uh, um, launch the futures based on the benchmarks, uh, you know, uh, talked about, um, in, in the associated contracts, um, per spec, per pending approval. Um, and then there's the physical matching venues, auctions, aggregators. So

[13:56](https://www.youtube.com/watch?v=GCse4OMzY8U&t=836s) anyone who can match capacity. We're looking at, um, essentially, uh, decentralized marketplaces for on-demand compute. You're looking at, um, liquid pools of compute that offer like liquid pool services and part it out themselves and have their own standardized contract. You're looking at bare metal resellers. You're looking at neoclouds, um, themselves doing larger capacity, uh, sales. And then you're looking at individual brokers and individual operators here as well on the bare metal reselling. Um, you have desks and brokers. They're l-looking at the bilateral quotes and execution. They might know a lot of people.

[14:26](https://www.youtube.com/watch?v=GCse4OMzY8U&t=866s) Um, they might be able to move fast. They might be able to see quotes at very similar specs in one place and know that there's a buyer that is, uh, extremely, uh, inelastic to price right now, needs the compute immediately, and they can get the RFS in time, so they can charge a decent spread on it. And then what they do is they're gonna have the source negotiating, and then they're gonna be carrying and placing the risk. Um, they're gonna have to operate that cluster for the person and-- but it has to be worth that spread. Um, and you know, you have looks in financial markets, uh, the exchanges that are coming, uh, that are based on the GPU hour, the futures and derivatives. Um, people are

[14:57](https://www.youtube.com/watch?v=GCse4OMzY8U&t=897s) gonna be using it to take a look at how to maybe hedge out and look into more, you know, financial structures. So hedge out risk. So hedge out, you know, my balance sheet risk, maybe get into more interesting, um, agreements. We're looking-- we're seeing kind of agreements such as, uh, prepayment payment facilities are coming online. We're seeing people utilize futures and-- or are interested in utilizing futures and other kind of, um, um, uh, options to be able to, um... There's demand for this to be able to hedge out the risk that they currently have on the books. We have buy-side

[15:26](https://www.youtube.com/watch?v=GCse4OMzY8U&t=926s) individuals, uh, looking for flex contracts, flexible contracts like, "Hey, is there ability that if I lock myself into a three to five year contract that I would be able to, you know, at least upgrade when a new chipset comes out? Or maybe I'll be able to roll it over. Maybe let me sell it back to you." And then there's like credit and financing, so equipment lenders, leasers, um, specialist credit providers. And so the way we're seeing it right now is that every deal is bespoke. There's different spec sheets. There's formats. The bill of materials may not match, um, the actual like delivery, or there

[15:56](https://www.youtube.com/watch?v=GCse4OMzY8U&t=956s) might be subject to change if a part is missing or a part is out of stock. And, you know, there's like very long lead times, um, that some of these OEM and ODM kind of, uh, providers have. And the same term can mean different things when you're negotiating very quickly. If someone's talking about memory, um, are we talking about, you know, uh, VRAM? Are we talking about the CPU memory? Are we talking about additional storage here? Like, what did you exactly mean there? It affects quite a bit what type of loads can be run and, you know, the actual end user's, um, needs, and also with pricing and delivery

[16:26](https://www.youtube.com/watch?v=GCse4OMzY8U&t=986s) dates. Um, so there's a lot of back and forth before anyone can really compare, and this is where the concept of like non-fungibility kind of comes into place and how to partition that out, what parts can be fungible, what parts hold the most of the weight, such as the SKU, et cetera, and what parts are still like traded as if it was a product, e-commerce, et cetera, where, you know, it's only for that particular cluster that this price applies.

[16:48](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1008s) And so prices can update daily or faster. Quotes still a good starting point. Bids can get shopped. Credit wording as changes what's on offer. There's all these kind of latent factors, hidden factors as well, why we're calling it a dark market. Um, these are some public H100 references snapshotted, um, today. I updated them. So take a look at the-- what the public-- So those are the references for if you wanted to go on their websites to see. Um, this is just an example here, um, of like what a public on-demand or spot pricing looks like for an H100, um, today. You can see kind

[17:18](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1038s) of various, various prices here. And so people might use this, or m-they might use their own kind of internal, what they've seen, um, to actually like assume or, or use it to charge what they want to charge on their actual cluster specification, take into account every other kind of factor that goes into it.

[17:36](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1056s) And so why are these new structures kind of needed, and what are these microstructures, um, interesting? Um, it's mainly to solve for these mismatches, right? Bridging capacity, more flexible commitments, resale rights, and protection against price moves. Um, you know, you're selling a fixed price service, uh, from an inference cloud point of view. You'd be selling a fixed price service while the GPU bill floats, and that creates price exposure. And so locking capacity for years can solve a part of that, but it'll create idle hour, uh, idle hour exposure elsewhere as well, unless you're always using it. Reselling spare capacity only helps if someone

[18:06](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1086s) actually can buy the hours. Uh, a financial hedge can address a reference price but can't deliver the cluster right now in terms of cash settlement, um, or guarantee a utilization. Um, so the useful structure has to start with the particular risk that the business has, um, and needs to manage.

[18:24](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1104s) And I'll, I'll go pretty fast on this policy section because we're running out of time. Um, everything that Affaan has said sits inside of a structure that is regulated and is affected by policy in various ways. One way is export controls. One of the examples, in January, the U.S. Commerce Department moved qualifying H200 class license reviews for China and Macau to case-by-case review with

[18:54](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1134s) conditions. Furthermore, open-weight models connect policy back to the U.S. economics we started with. The U.S. AI Action Plan explicitly encourages open source and open-weight AI and recommends improving the financial market for compute so that startups and academics can access large-scale computing power.

[19:17](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1157s) So I'm going to tie this out right here, but where does this opportunity lie? We have this growing middle market. We need balance sheets that can bridge buyer and supplier needs. There needs to be some level of standardization, a standard capacity contract, common specs, common terms, um, and definitions. And so growing companies whose needs and commitment do not fit the contracts that hyperscalers and frontier labs can sign, those companies have real workloads. We want them to grow. They still face this financing problem, a timing problem, a specification problem. A supplier might want the long-term commitment for the build-out while the buyer wants to expand in smaller

[19:47](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1187s) blocks as demand becomes clearer, as it's so volatile, front of curve. Uh, the buyer needs a way to avoi-avoid paying for capacity it can't use yet. And so participants with strong balance sheets that are uncorrelated with the market and the ability to assess those risks with a strong credit profile can help bridge the gap between what, um, buyers need and what suppliers offers. That requires more than quoting the actual GPU hour. It means understanding the delivery, workload, all of that. And so there's very few kind of players in the market and, you know, a very well-suited player, um, can be a firm that's uncorrelated with market, has a strong balance

[20:17](https://www.youtube.com/watch?v=GCse4OMzY8U&t=1217s) sheet, strong credit rating, and incredible amounts of hardware talent that sits inside for high-performance computing. And so just making a CTA here, let's build this together. Thank you guys so much for attending our talk.
