Stepan Braginskiy

Financial logistics for AI compute infra

[Stepan Braginskiy @StepanVC]

Wouldn’t a much simpler demand response program be a dynamic, preemptible spot price like the one Nebius launched?

H200 prices fluctuate from $5.4/hr to $0.79/hr off-peak in a day. Consumers can ask an agent to delay running workloads (that aren’t time-sensitive) until the price is less than X. No need to pre-sell capacity.  No need to invent credit-conversion AMEX-like infra. 

What do you think?

[Stepan Braginskiy @StepanVC]

At Jev's release on September 15, @typesafeai was valued at $200 million in its seed round.

Now, according to the Financial Times, investors are offering funding that would value the company at $10 billion.

If you compare prices, Jev's "System One" decision model is 95 times cheaper per input token than Opus 5.5. Output tokens are free.

It didn't take long for competition to start eating into Anthropic's and OpenAI's crazy margins.

It just reminded me that OpenAI and Anthropic make $100M per 1MW of compute capacity in inference API revenue, compared to Neocloud's $12M/MW in 5-year average IaaS revenue, i.e., they are 8 times more expensive than running AI workloads on GPU bare metal, as SemiAnalysis reported in August.

That leading labs' pricing advantage will only last until developers and enterprises switch to more cost-efficient, use-case-based models like TypeSafe AI.

Will Anthropic's and OpenAI's IPOs hold up?

[Stepan Braginskiy @StepanVC]

Keep hearing that mid-market neoclouds face limited funding options and are on the lookout for equity.

It is true that sub-institutional capital has not yet fully arrived to match the mid-market financing demand.

But family offices and private credit will not pass on a good deal that yields more than comparable risk assets they invest in.

The bottleneck is: (1) distribution — accessing the right investors, and (2) structuring — offering a deal that clears IRR hurdles under stress assumptions.

No one signs a deal on assumptions that are hard to justify.

[Stepan Braginskiy @StepanVC]

The scarcest asset in mid-market AI infra isn't power or GPUs. It's an offtaker a lender will underwrite.

Last month Lambda closed a $926 million GPU loan rated investment grade. Lambda is a private company. The investment-grade name in that deal is the customer.

Hut 8, a former bitcoin miner, sold investment-grade bonds to build a data center. Google stands behind the rent.

Both borrowed on their customer's credit, not their own. The hardware is collateral. The offtake contract is what repays the loan.

Mid-market operators can buy the same GPUs and build the same site but often cannot get an offtaker of that quality. That raises the cost of debt from single digits to the mid-teens when it is available at all.

That makes the offtake the first problem to solve, not the last. Who is signing your offtake, and for how long?

[Stepan Braginskiy @StepanVC]

Thanks for the correction! Saw @sfcompute enter into multi-year offtake commitments. Congrats! I have a question - does this offtake sit in a book that your compute providers use to underwrite hardware debt for new builds, or does it only count as their revenue against existing capacity?

I'd really love to hear how you are solving the bankability of mid-market demand problem. Please DM if you are open to a quick chat.

[Stepan Braginskiy @StepanVC]

Mid-market access to AI compute is a credit problem

A few days ago, Nebius Group CEO Arkady Volozh noted that demand for AI compute is "unlimited", with customers reserving capacity into 2028. Yet Nebius's growth is constrained by how quickly it can build and how much construction it can finance.

The same story holds for almost every major neocloud today.

Neoclouds prioritize allocating capacity to frontier AI labs, large enterprises, and hyper-capitalized startups. These customers can afford multi-year, take-or-pay offtake contracts. More importantly, their balance sheets or backing (VC, strategic, sovereign) are strong enough for neocloud lenders to underwrite the underlying GPU debt, allowing neoclouds to bring more capacity online.

This leaves a long tail of demand from early- to late-stage AI-native startups severely squeezed.

The immediate result has been the rise of compute marketplaces like @sfcompute , @vast_ai , @runpod , and others. They scoop up idle compute capacity from the market and resell it in smaller increments at a margin. This helps developers and startups cover their compute needs for experimentation and validation.

But marketplaces offer little for AI natives graduating their workloads to production. These mid-market buyers need reliable, multi-node GPU clusters at a cost that makes unit economics work. This capacity is rarely available on demand.

Marketplaces partially solve the allocation problem. They do not solve the capacity problem. And capacity is a credit problem.

Mid-market buyers have sizable AI compute needs and improving credit profiles. They can afford offtake contracts spanning months to a year. Individually, however, their standalone offtake size and risk profile are unlikely to qualify for the project-finance debt required to underwrite GPUs. Consequently, operators allocate primary capacity to those who qualify.

The solution requires a capital markets primitive: demand pooling.

When aggregated, a single master contract qualifies not just by total volume, but by an improved credit profile. Pooling introduces diversification, reducing unsystematic default risk for infrastructure lenders.

That turns mid-market demand into bankable borrowing capacity. Operators get the offtakers they need to finance new GPU hardware, and mid-market buyers get the primary AI compute capacity they are locked out of today.

If you are scaling AI workloads to production right now, how are you balancing the premium of on-demand compute against the rigid balance-sheet liabilities of long-term offtakes?

[Stani Korostelev @AmusingVentures]

I counted the companies building market structure for AI compute: indices, venues, desks, exchanges, credit and insurance around the GPU-hour, not the clouds that sell it. I got to 58 in seven boxes, starting from the @socialgraphvc primer and the @0xfishylosopher essay, and I am fairly sure the real number is higher.

Price discovery. @silicondata runs the H100 and B200 rental indices that CME's October futures are designed to settle on, @OrnnExchange publishes OCPI and a volatility index and has ICE as its futures partner, @ComputeDesk puts IOSCO-compliant indexes built from private OTC transactions on Bloomberg, NATIVX normalises GPU prices to an energy unit, and @SemiAnalysis_ tracks contract prices and runs the ClusterMAX census of 209 GPU clouds.

Spot venues, auctions and aggregators. @sfcompute runs an order book where contracts can be resold, @computeexchange auctions reserved capacity for one to 36 months and opened a used-GPU secondary market in July, @getcomputable sells dedicated H100 nodes by the calendar week in sealed-bid auctions, and @stoaexchange discovers clearing levels for the hardware itself. @mithrilcompute, @shadeformai and @get_hydrahost aggregate other people's clouds into one buying surface, and Hydra raised $100M in June for that, with an offtake network attached. NVIDIA's DGX Cloud Lepton does the same from the top of the stack, while RunPod, https://t.co/J8NJctvoBA, TensorDock, Salad and Clore run the spot layer.

Decentralised networks. @akashnet, @ionet, @AethirCloud, @rendernetwork, @nosana_ai and @spheron match GPU supply and demand with on-chain settlement, @fluence_project started auctioning reserved GPU clusters in August with 6,000 GPUs across 13 countries, @PrimeIntellect and @GensynAI come at the same market from distributed training, and Hyperbolic, Golem, Theta EdgeCloud, Aleph Cloud, Lium and Targon fill out the box. A large share of this whole map is crypto-native, which the primer also noticed.

Desks and brokers. @itomarkets quotes a single fixed GPU-hour rate with an SLA and stands as principal, @intheanera runs a forward capacity market for deliverable inference and calls itself a clearinghouse for AI risk, and the incumbents have arrived: @BGCGroupInc opened an OTC desk for secondary compute and memory capacity in June, and @FalconXGlobal dealt what it calls the first OTC swap on the forward price of compute, referencing Ornn's H100 index.

Exchanges and derivatives. @CMEGroup, @ICE_Markets and @NodalExchange have all announced GPU futures this year, Nodal on Compute Desk's indexes next to its power contracts. @Architect_Fi builds GPU and DRAM perps with a bridge into physical delivery, @liquidcompute is going the CFTC exchange and clearinghouse route, @MNX_fi runs H100 perps and lab valuation futures, Hyperliquid already trades H100 perps, Kalshi and Polymarket run GPU price markets, and @castle_tech_ builds markets for the exposures a data center cannot hedge today, permits first.

Credit. @USDai_Official lends up to 70% LTV against tokenised warehouse receipts on the hardware, @gaib_ai and @Compute_Labs turn GPU financing and GPU ownership into yield-bearing tokens, @cysic_xyz does the same for mixed hardware, and @gpufinancing writes plain GPU equipment loans and leases. More than $20B of GPU-collateralised loans are outstanding, and in my experience the borrowers below investment grade still mostly sit outside that number.

Insurance, collateral and plumbing. @amcompute1 writes declining residual value floors for up to three years, @PRINCEPSdev turns SLA terms into coverage carriers can bind, Barkr issues Munich Re-backed warrantied valuations of GPU collateral, @Aravolta_X25 monitors the collateral through telemetry, and @netbackyard does the metering and billing that turns capacity into a tradable unit.

And the incumbents moved faster than I expected: CME, ICE and Nodal announced GPU futures between May and September, and BGC opened its desk in June.

If you are building in any of these boxes and you are not on the map, reply with what you do.

If I put you in the wrong box, pls say so.

There is also a WhatsApp group for people building in compute markets, so DM me or @jessiedong_ if you want to be added.

[Stepan Braginskiy @StepanVC]

Would agree. Offtakers must be credit-scored to ensure the pool has acceptable default risk. Operator credibility also matters because execution risk is high, even if they bring reliable offtake to a lender.

Interesting that usd​.ai admits first-time operators on paper but enforces a liquidation trigger on the GPUs if the operator misses a payment by 30 days. Why would any first-loss equity sponsor accept that structure?

Analogy: a retail investor trading on leverage, getting wiped out by a margin call.

[Stepan Braginskiy @StepanVC]

I'm 40.

VC turned founder in Tokyo.

Looking to connect with people building in AI infra and compute markets.

[Hauke 🌱🦀 @haukejung]

I'm 38.

Solo founder from Germany.

Looking to connect with more builders & indie hackers!