A single number crossed my terminal: 100MW.
Solowin will reach 100 megawatts of AI compute capacity by 2028. That was the payload. There was no site address, no power purchase agreement, no interconnect queue position, no GPU allocation letter, no named customer, no capital structure.
I have spent the last decade auditing supply claims. In 2017, I burned six weeks inside the EVM bytecode of a privacy coin called Project Aether and found a minting function the whitepaper never disclosed โ a 12,000 ETH discrepancy between stated and actual supply. The project delisted from three exchanges within a week. The governing lesson from that engagement has shaped everything I have published since: a number is not evidence. A number is a hypothesis wearing a suit.
Solowin's target was reported by Crypto Briefing, bundled with a second observation โ that the U.S. Congress remains stalled on crypto legislation. Those two facts were packaged together, and the packaging is itself informative.
For a claim like this to become an asset rather than a press release, it needs four verifiable components. First, an enumerated site with a power delivery date. Second, a signed power purchase agreement with a named utility, or a disclosed interconnect queue position. Third, a GPU procurement contract with allocation confirmed by the vendor. Fourth, at least one offtake agreement with a customer who has committed dollars. Any AI infrastructure operator with real economics can produce at least three of those four inside a data room.
Solowin has produced none of them publicly.
The demand side is not in question. NVIDIA's data center revenue trajectory between FY2023 and FY2024 reflects genuine, contracted pull. The question is never whether demand exists. It is whether a specific operator can convert that demand into invoices โ and at what cost of capital.
Start with the unit of account, because this is where most readers get lost.
A megawatt is an input, not a product. You cannot invoice a megawatt.
100MW is facility load. Apply a realistic power usage effectiveness of 1.3 for a liquid-cooled AI hall and you are left with roughly 77MW of IT load. Divide by the 700W thermal design power of an H100 and you get approximately 110,000 accelerators. At 2024 street pricing of $25,000 to $30,000 per unit, the silicon alone lands between $2.8 billion and $3.3 billion. The building โ the part everyone fixates on โ is the cheap component, typically $150 million to $300 million.
Data indicates that Solowin disclosed the cheaper half of its own balance sheet and omitted the expensive half.
Now place the number in the competitive set. CoreWeave operates on the order of a gigawatt of contracted capacity. Microsoft, Google, and Meta collectively committed more than $200 billion of capital expenditure in 2024 alone. A 100MW facility is roughly a tenth of one established AI-cloud peer and a rounding error against hyperscaler budgets. The claim that this expansion "reshapes the competitive landscape" is not supportable by any disclosed figure.
The structural choice matters more than the size. Render and Akash do not own megawatts; they broker idle capacity through token incentives, which produces a light balance sheet and a weak moat. Solowin has taken the opposite path โ heavy capex, heavy depreciation, heavy financing risk. That is a legitimate strategy. It is also the strategy that fails hardest when rates stay elevated or GPU rental pricing compresses. In a bear market, this is the risk parameter that matters: not whether the target is achievable, but whether the financing survives the three years before the first invoice.
The missing ledger is the finding. Here is what would move this from narrative to thesis:
| Required disclosure | Status | Why it matters | |---|---|---| | Site location and power delivery date | Undisclosed | Determines real construction timeline | | Utility PPA or interconnect queue position | Undisclosed | Transformer lead times run 18โ36 months | | GPU allocation confirmation | Undisclosed | Allocation is the scarce asset, not land | | Named offtake customer | Undisclosed | Converts capacity into revenue | | Funding structure and equity committed | Undisclosed | $3B+ exposure requires disclosure |
One layer deserves a flag rather than a conclusion. The U.S. legislative pipeline for digital assets โ stablecoin frameworks, SAB 121 treatment, market structure bills โ has moved slowly, leaving mixed AI-and-crypto operators in a zone of strategic ambiguity. Positioning a facility as "AI compute" rather than "crypto mining" changes the regulatory perimeter it sits inside. I cannot verify intent here, and I will not claim it. I will note only that the framing choice is available, and that it is cheap.
The consensus reading of this announcement is straightforward: AI plus crypto crossover, bullish.
The counter-reading is narrower and more uncomfortable. Megawatt targets are the cheapest claim in this industry to issue and the hardest to falsify in the near term. They require no contract, no counterparty, and no capital to announce. Over the past eighteen months they have increasingly functioned as a financing instrument โ a headline that precedes a raise, rather than a description of a build.
This is where correlation discipline matters. AI compute demand is growing. That fact is true and verifiable. It tells you nothing about which operator captures the demand, at what margin, or on what timeline. Demand growth and operator viability are separate variables, and the second one has a far higher mortality rate.
I have seen this sequencing before. In 2022, monitoring Terra's reserve addresses, I watched collateral quality deteriorate roughly 40% three days before any public statement. The data preceded the headline. Chain links don't lie โ they simply arrive earlier than press releases.
Here the situation is inverted. There is no wallet, no contract, no transaction, no reserve address. Code is the only witness, and in this case there is nothing for the code to witness. That absence is the analytical output. Wallets connect the dots; when there are no dots, you learn something about the shape of the claim.
Watch four signals over the next two quarters. An interconnect queue entry or an executed PPA. A transformer procurement order with a named supplier. A GPU allocation confirmation from NVIDIA or an authorized channel partner. And a single signed offtake contract with dollars attached.
If none materialize, the 100MW number was never a plan. It was marketing. Follow the gas, not the hype.