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. @Silicon_Data 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, Vast.ai, 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.

OpenAI = Walmart Supercenter: one-stop shop, huge scale. Anthropic = Walgreens / CVS: still general retail, but a bit more curated. TypeSafe / Jev = Costco: warehouse-club prices.

but 60% of these <6 ppl teams manage aum <$500m. the complexity of family office operations grows significantly in the billions range. in my family office experience we were 40 people managing $6 billion AUM, and we were actively involved in asset management, with in-house experts in metals&mining, project management, audit, security, etc. can't imagine we could manage it with 5-10 people.

The lender behind Nscale's Norway AI data center is the same one that finances Norwegian trawlers and oil rigs. Eksfin, Norway's state export-credit agency, joined ABN AMRO, DNB, Nordea and SEB in a $790M facility for Nscale's Narvik campus, plus a further $790M accordion for expansion. Export credit exists to back anything built with Norwegian power and labor, and an AI data center with Microsoft as the offtaker for 30,000 Nvidia GPUs there now qualifies. That's a different underwriting logic than the US deals we've seen, where debt gets priced off whoever signs the offtake. Eksfin isn't underwriting a customer. It's underwriting industrial policy. That's a financing channel the US market doesn't have (right?). An operator building where a government wants to attract AI investment should be asking whether their export-credit agency does this too.

One day we’ll see InDrive for compute, so users will be able to offer and negotiate prices.

I mapped the listed vehicles I could find that put public/retail money into GPUs and data centers. I got to 35 names in seven boxes, plus one empty box. SPACs and shells. @BoostrunGPUs, a bare-metal Nvidia cloud, merged with Willow Lane in May, got the full $134.5M trust because nobody redeemed, and reported $31.1M of revenue last quarter with a $1.44B Dell purchase agreement behind it. Exascale Labs took the same route with D. Boral's $250M SPAC in August and came out with about $12M of cash, and the stock now trades near $2.80. @sharon__ai merged into a Roth SPAC and traded on OTC in December, then raised a $125M Nasdaq IPO led by Oaktree in February and $1.05B of converts since. @GoodVisionAI votes on its $180M merger with Calisa in October, @TECfusions is merging into Apex Treasury's $345M trust at a $4.2B enterprise value, and Host Digital reverse-merged into a wellness company last week with a 43 MW Oklahoma site and $1.25B of contracted revenue. AI Infrastructure Acquisition Corp still holds $138M on the NYSE with no target, and Silicon Valley Acquisition II filed for $220M this week with AI infrastructure on its list. IPOs. @WhiteFiber_ raised $159M on Nasdaq in August 2025 and $310M of 5% converts a year later, @FermiAmerica raised $682.5M as a REIT for its Amarillo power campus and signed TensorWave for 222 MW in August, and @nscale filed for a NYSE listing last week after a Series C at $14.6B. Former bitcoin miners. Most of the listed capacity sits here, because these companies already had power, land and a ticker. @IREN_Ltd sold $2.6B of 1% converts in May against a $9.7B Microsoft contract. @terawulfinc, @CipherInc and @Hut8Corp leased sites to Fluidstack with Google backstopping the rent in exchange for warrants, and Hut 8's River Bend lease alone is $7B over 15 years. @Core_Scientific voted down CoreWeave's takeover and kept about 590 MW of CoreWeave hosting, @APLDDigital is at $31B of contracted base lease revenue, and @GalaxyHQ delivered the first 133 MW of Helios to CoreWeave. @Bitdeer, @BitDigital_BTBT, @HIVEDigitalTech, @SolunaHoldings and @hyperscaledata are making the same turn at smaller scale, Hyperscale Data by borrowing about $30M against its bitcoin on Morpho at 4.9% to build in Michigan. Outside the US. @sakura_pr runs a GPU cloud in Hokkaido with up to ¥50.1B of METI subsidy, and @NorthernDataGrp is now 85% owned by Rumble and leaving the Munich exchange. Listed funds. Cordiant Digital Infrastructure is the closest thing to what I had in mind: a £965M London-listed closed-ended fund whose Czech business sells GPU-as-a-service and is building a 26 MW data centre in Prague, trading about 14% below NAV. Infratil in New Zealand owns 49.7% of CDC and 55% of Kao Data, and Digital 9 shows the other end, in wind-down with NAV at 8.6p a share. ETFs. @Grayscale renamed its Bitcoin Miners ETF to the AI Compute ETF ($GCPU) yesterday, same fund with $12.8M in it, and its index now picks companies for their AI and HPC work rather than mining. @roundhill launched a Neocloud ETF in August that reaches Nebius, CoreWeave and IREN through swaps, and @GlobalXETFs, @vaneck_us and @iShares run data centre baskets. None of them can buy a GPU, because a US fund can hold at most 15% in illiquid assets, so this route only ever buys the listed shares above. On-chain. @USDai_Official took a $100M facility from Bullish in August to lend against GPUs, and @gaib_ai runs AID and sAID with about $19M locked. They lend against the hardware directly, but they are not listed securities (yet). The box I could not fill is a small listed vehicle that lends against GPUs. Hercules and Trinity are listed tech lenders and I could not find GPU-backed loans in their disclosures, so today the only people doing that for mid-market clusters sit on-chain or in private credit. -- If you run or are building a listed vehicle for compute and you are not on the map, reply with what you do. If I put you in the wrong box, pls say so. If you are in the compute business and want to join the niche WA group with founders and researchers, DM.

When a family office diligences an AI infra deal, it tries to answer (at least) 4 questions: Who pays? They underwrite the offtake, not the GPUs. The tenant (or depth of your pool of on-demand clients) largely sets the equity return. What's the structure? Equity is 30-50% of equipment cost, and the lender is the remainder with a minimum DSCR. Who runs it? A family office backs the operator before the specific project. How do they get out? Lenders rarely want to go past 36 months, so say up front whether the exit is a refinancing or a sale of the SPV or hardware, and what residual value you assumed.

Floating data centers are four different products: One: a barge at a quay that uses the river as a cooling loop and takes power from shore. @NautilusDT ran it at Stockton from 2021 at 6.5 MW, then listed the barge for $45M at 86% leased and now sells the cooling units instead. MOL and @Hitachi want to convert a second-hand ship by 2027. Keppel is the exception: 25 MW under construction in Singapore with a hyperscaler already signed, because there the constraint is land, not water. Two: a shipyard-built hall moored beside a power plant the grid queue cannot deliver. Samsung Heavy and M3 are engineering 50 MW units next to a gas plant in Houston. Three: fully offshore with your own power ship. Atomarine (@DimitrisKoute) wants 75-100 MW barges 10 miles out, gas ship first, reactor ship later, and a reactor-barge power layer is forming underneath: Bluecore ($50M seed), Core Power, Saltfoss, Deployable Energy with Hornbeck's supply vessels. Four: open ocean on wave power and satellite. @_panthalassa's 85 m spheres, Mocean, NetworkOcean, all aimed at inference. I 100% understand these projects represent interesting technical challenges (TRIZ at its best), but what are the exact demand categories? Is it only military cases requiring autonomy/low latency?

Venture debt could be part of the answer to this mid-market compute gap. Venture lenders already underwrite cash-burning startups. The question is whether the buyer can support the debt and the compute commitment, not whether it has positive EBITDA today. From my recent conversation with a partner at a top-10 European venture debt platform: - runway starts with cash divided by monthly burn. An older shortcut used average EBITDA losses over 3 months, but now they use average monthly operating and investing cash burn over 6 months. - for conventional venture debt, his preference was 9-12 months of runway, ideally 12. He also wants monthly P&L, balance sheet, and cash-flow reporting, revenue growth, and the amount, timing, and quality of the last equity raise. The next funding milestone and existing investors' appetite to support it matter A LOT. Applied to @distributionat's post, I would test runway after the deposit and include ongoing compute bills and debt service in the cash forecast.

If one appointment ruins the day I see two reasons: (a) you don’t like what you do in business/job/role and/or (b) you have issues with planning - meetings can be packed back-to-back and weekdays can be split between talking and deep work (zero meetings)

The initial energy buildout thesis - batteries and solar panels - was appealing narrative as new source of DeFi yield. I remember the post from @StaniKulechov on Aave’s vision on opportunity to finance it. With maturing AI compute markets, better variety in AI monetization business models, and real pull from the neocloud operators on finding non-dilutive financing we are finally getting very close to real infrafi.

Can you explain this 50% GPU capacity thing? I assume there are two different utilization numbers: One is physical: how much of the time a GPU is actually computing versus idle, waiting for data over the network, between GPUs, and between racks. With a decent topology, I would expect that number to be nowhere near 50%, but okay. The other is financial: how much of a contract's paid capacity a customer is actually using. This is the one lenders watch. The rule of thumb we hear is close to 80% before a bank can call the loan and take the collateral. Meaning a cluster can be network-bound and still be bankable.

It caught our eye that dealers are asking 4.09x what used DDR4 32GB RDIMMs actually sell for. This is list prices vs cleared prices. The typical sale price is $100, while the dealer ask is $410. Used DDR5 64GB RDIMMs sell for $1,300 against $3,000 asks. With new supply on allocation and DDR4 out of production, dealers price used memory listings off scarce new modules. GPUs sit much closer, with used A100 80GB GPUs selling for $5,200 against $6,228 asks. Interested in learning more? Message me or check out our Rack Report where this came from. Download it here: amcompute dot com

Sports franchises ran at 20-30% loan-to-value for decades and looked underleveraged. It turned out to be one of the most durable asset classes. @ApolloGlobal's Jim Zelter drew that line on @bsurveillance this morning and pointed it straight at AI buildout financing, low leverage next to selective equity in assets built to survive disruption, the same split Apollo just did with the Yankees and Atlantic Aviation. Interesting in the context of the equity/debt topic for AI infra.

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?

Interesting take on mezzanine in a GPU financing. Let's say we cut a slice of equity and debt and replace it with $7.5M of mezz in a 1 MW cluster with $35M in initial debt (36 months) and $15M in sponsor equity. The bank usually lends only on condition that all project cash goes to repaying the bank first, while the mezz is normally allowed to collect interest during the term, but not principal until the bank is repaid or well ahead. Hence the mezz is mostly lump at the end. So, in a very basic illustrative scenario, by month 12 the debt is roughly 1/3 ($23M), and the remaining $7.5M mezz is covered in full only if the cluster still has residual value of 62% of cost ($31M/$50M), and is kind of fully wiped at 47%. Therefore, the GPUs' value decay rate is super important here. Whoever writes it is holding a used-hardware position and should underwrite the resale bid before the borrower. Am i right?

I’m 37. Founder from Munich. Looking to connect with people building in AI infra and compute markets.

It seems mid-market-sized offtakers are the inevitable next challenge for AI infra financing. @BrettHarrison expects CDOs on GPU-backed leases soon, because operators below investment grade cannot borrow at double-digit rates alone. That only works if the credits inside the pool actually differ. Pool them carelessly, and you have stacked the same GPU bet, not diversified it, close to what @wayne_nelmz called the 2008 mistake: banks marked mortgage bonds to a model instead of a price and only learned the value when forced to sell. Before anyone pools these credits, someone needs to score which offtakes are actually creditworthy, but with AI companies running on 20-30% margins, it's more demanding than in the revenue-based credits during the SaaS era.