The Race to AI Is Now a Race to Power — and Sam Altman Is Burning Gas to Win It
What you’re reading: a real Opportunity Dossier from our AI research pipeline — Ada spotted the signals, Bea shaped the candidates, Cora dug deep, Edie tried to kill it and owns the score, and Della wrote it up. New to our scoring? Start with How Our Star Score Works. Research date: July 2026 — facts move; check dates before betting money.
Fluidity Rating: ★☆☆☆☆ (1/5) — a genuinely hot market with a genuinely bad default startup: the accessible version is low-moat and slow to sell, and the version with a real moat needs the kind of power access, permits, and financing that already sit with giants.
A grid-software founder named Seyed Madaeni says the whole game has changed shape. Sam Altman, the most famous face in artificial intelligence — the company he runs, OpenAI, is building power-plant-sized computer campuses — admits, standing in the Texas heat, exactly how the lights stay on. Chase Lochmiller, who started a company scavenging waste gas from oil fields and turned it into a computing empire, keeps hammering one idea: build around the electricity first.
Here is Madaeni, chief executive and co-founder of Verse — a company that builds software to help data-center developers juggle on-site power — in the letter announcing his June 18, 2026 funding round: “The race to AI is now a race to power, and developers are losing time they don’t have.” (verse.inc)
Here is OpenAI, in an April 29, 2026 post, describing its “Stargate” mega-project as something that “brings together companies with deep expertise in cloud infrastructure, data centers, chips, energy, construction, finance, and operations,” and says OpenAI is already planning beyond an initial 10-gigawatt goal — 10 gigawatts being roughly the output of ten large nuclear reactors, an almost cartoonish amount of electricity for computers. (openai.com)
And here is the tension that makes the whole story honest. At OpenAI’s Stargate site in Abilene, Texas, the Associated Press reported that Sam Altman said plainly: “We’re burning gas to run this data center.” (apnews.com)
That collision — grand clean-energy ambition on one side, a gas turbine roaring right now on the other — is the entire opportunity in one breath. OpenAI talks in gigawatts and clean futures; OpenAI runs on gas today because gas is what can be energized today. Not hypocrisy — the raw compromise every player in this field is making between the power mix they want and the power they can actually plug in this quarter.
What world is this, and why should a stranger care?
Let us make the opportunity plain. Distributed compute and energy-integrated data centers for AI = the business of matching hungry AI computers to electricity faster — either by building smaller, movable computer units next to spare power, or by selling software and services that help developers skip the years-long wait to get plugged into the grid. The computers that run AI eat staggering amounts of electricity, and the bottleneck is no longer chips or land — it is energized, permitted, financeable megawatts. Getting power to a site can take years. Everyone in AI is now scrambling for a shortcut.
Why should you, a stranger, care? Because the money is real and enormous. The International Energy Agency reported that five large tech companies’ capital spending topped $400 billion in 2025 and is set to rise another 75% in 2026, much of it chasing exactly this power-and-compute problem. (iea.org)
The three things you’ll remember:
- The market is a five-star fire, and the startup built on top of it is a one-star trap — hot problems don’t make good founder bets.
- The moat lives in power, permits, and utility handshakes — none of which a coder can type into existence — the accessible version is the least defensible one.
- The giants already own the obvious ground — Crusoe at a $10B-plus valuation, CoreWeave with $2 billion from NVIDIA, Applied Digital with a $5 billion financing line — so a generic “energy-integrated AI infra” pitch shows up late.
The allure — and it is a real allure
Let us sell the dream honestly first, because the dream is not fake. The demand here is not a projection dressed up as a trend; it is measured, and it is steep. Berkeley Lab reported that U.S. data-center electricity use climbed from 58 terawatt-hours in 2014 to 176 terawatt-hours in 2023 — a terawatt-hour being roughly the yearly electricity use of about 90,000 U.S. homes — and could reach 325 to 580 terawatt-hours by 2028. (energyanalysis.lbl.gov)
The federal Department of Energy, citing the research group EPRI, says data centers could consume up to 9% of all U.S. electricity generation each year by 2030, up from about 4% in 2023. (energy.gov)
And the crunch is visible in the real estate. The commercial-property firm CBRE reported that data-center vacancy across North America fell to just 1.4% in the second half of 2025 even as total capacity grew to 9,432 megawatts — a market so tight that anything you can energize, you can rent. (cbre.com)
For an investor, the lens is simple and thrilling: whoever can turn stuck power into working AI capacity, months faster than the grid allows, is selling the scarcest thing in the fastest-growing corner of the economy. The believers argue the value doesn’t all flow to the campus owners. Some of it, they say, belongs to whoever builds the software layer — the tools that help a developer squeeze more out of on-site batteries and generators, jump the interconnection queue, and turn idle behind-the-meter power into revenue. That layer is capital-light, high-margin, and — the pitch goes — still up for grabs.
Where this idea came from
The provenance runs through real people, and the most emblematic is Chase Lochmiller of Crusoe. Crusoe began as a genuinely strange idea: capture the waste methane that oil wells flare off into the sky and burn it to power mobile, modular data centers parked right at the wellhead. As AI demand exploded, Crusoe pivoted that “start with the energy” instinct into building full AI factories. Crusoe announced on October 24, 2025 an initial close of a $1.375 billion Series E funding round, saying it brought the company’s expected valuation to over $10 billion. (crusoe.ai)
Then OpenAI made the gigawatt conversation mainstream. Once the most-watched company in AI started talking publicly about campuses, energy composition, and construction financing, “how will you power it” became a founder-imagination question, not just a utility-planning one. The April 2026 Stargate post cemented that language. (openai.com)
The countermove — the “distributed” half of this opportunity — emerged as interconnection queues hardened into years-long walls. Newer players began pitching that many AI workloads, especially inference (running an already-trained model to answer requests, as opposed to the massive job of training one), don’t need to live in a single billion-watt fortress. Armada, for one, said in a July 2025 announcement that it raised $131 million and described its “Leviathan” product as megawatt-scale modular data centers that can sit beside stranded gas, solar, nuclear, or other power sources. (armada.ai)
Who wants to build it
Here is the honest signal, and it is a warning. The strongest activity is not coming from garage-stage founders. It is coming from company combinations and heavily capitalized players. Verse itself announced a $54 million Series B on June 18, 2026, led by the venture firm Bessemer Venture Partners with participation from GV, NVIDIA, and Norrsken. (verse.inc)
There are faint signals of an earlier-stage founder wave forming. The Chicago hardware incubator mHUB launched a “compute-energy nexus” accelerator in May 2026, suggesting institutional belief that a startup cohort is gathering around this exact intersection — though that accelerator is a single-source signal, not yet corroborated proof of a broad founder rush. Our research found isolated founder chatter on forums but no visible garage-founder wave comparable to AI agents or developer tools. In plain terms: the people winning here arrive with balance sheets and utility relationships, not with a clever prototype.
The Scorecard — nine measures of Return on Founder Time
Return on Founder Time asks a blunt question: for the months and dollars a founder pours in, what comes back? Here is where each of the nine measures lands for a cold stranger with no power-industry Rolodex.
Time-to-Truth — ★☆☆☆☆. How long until you honestly know whether to keep going? For the only sane wedge — software and orchestration for faster power-to-compute — a trustworthy buy/no-buy still takes roughly 4 to 9 months and $150,000 to $750,000, because you need real counterparties, utility-adjacent diligence, and one to three live pilots before anyone trusts you with revenue-critical power decisions. Touch owned compute or sites and truth stretches to 9–18 months and far more cash.
Time-to-MVP — ★☆☆☆☆. A credible minimum product is not a weekend app. Expect 3 to 6 months with a 4-to-8-person team for the asset-light version, including integrations with the industrial control and battery-management systems that run power sites. A version that owns modular compute or a site is more like 9 to 18 months plus hardware and permitting.
Time-to-First-Customer — ★☆☆☆☆. First paying customer is likely 6 to 18 months from a cold start, because the buyers — data-center developers, large-load project developers, independent power producers, and enterprise infrastructure operators — run long trust cycles and bespoke procurement. A founder with existing industry relationships might compress that to 3 to 9 months; a stranger will not.
Capital Requirements — ★☆☆☆☆. For the only startup-shaped wedge, founder cash is likely $250,000 to $1.5 million and total capital $1 million to $5 million to fund a serious team, integrations, legal work, and pilots. Any version that owns modular compute, power assets, or sites jumps quickly into $10 million to $100 million-plus territory. This sits far above normal bootstrapped range.
Execution Complexity / Operational Risk — ☆☆☆☆☆ (zero). This is a dependency swamp: utilities, interconnection realities, project finance, energy contracting, uptime liability, hardware-software integration, custom engineering, and permitting exposure. Even the asset-light version has many failure points; a lightly staffed founder is very likely to get buried. This zero is what pins the whole rating near the floor.
Competitive Intensity — ★☆☆☆☆. The obvious narrative is already crowded by giants. On January 26, 2026, NVIDIA and CoreWeave announced NVIDIA’s $2 billion stock investment and CoreWeave’s plan to accelerate buildout of more than 5 gigawatts of AI factories by 2030. (sec.gov)
And the financing muscle behind the campus-owner version is on another planet from a startup. Applied Digital announced a financing facility of up to $5.0 billion with the asset manager Macquarie to support its high-performance computing and data-center growth. (sahmcapital.com)
Market Momentum — ★★★★★. The lone five-star pillar, and it is fully earned. Demand is real, measured, and rising fast, corroborated by Berkeley Lab, the Department of Energy citing EPRI, and CBRE’s record-low vacancy. The bottleneck is genuine — even though this particular startup variant is weak.
AI Leverage — ★★☆☆☆. AI helps with optimization, dispatch, siting analysis, and forecasting, but it does not remove the hard parts: relationships, interconnection, permits, hardware, financing, and uptime liability. Worse, AI tends to commoditize generic orchestration software unless that software is deeply embedded in real project delivery.
Defensibility — ★☆☆☆☆. The strong moats — secured power, land, permits, utility relationships, financing capacity, delivery reputation — are exactly the assets a new founder does not have. Pure software or brokerage is easy to start and easy to copy unless it becomes the system-of-record inside real projects. So the accessible version is the least defensible one.
How customer growth compounds: weak. Customer number two is not meaningfully cheaper to win because customer number one exists — unless the company becomes a trusted industry node earning repeatable referrals among developers, engineering-and-construction firms, utilities, and power-asset owners. Mostly this is relationship sales and bespoke delivery, not a self-feeding distribution engine.
Cost per completed outcome: for the direct-sales motion that fits, roughly $1,000 to $5,000 in founder time, travel, engineering support, and diligence per serious opportunity — not per closed deal, per genuine shot at one.
What It Would Take to Raise This Score
Your money. A founder would personally invest or commit roughly $250,000 to $1.5 million (via an angel-friendly bridge), while the business as a whole needs about $1 million to $5 million for the asset-light software-and-services wedge — and $10 million to $100 million-plus the moment it touches modular compute, generation, or sites. All of these figures are estimates, not measured facts.
Time to break even. An estimated 18 to 36 months to cover costs for the asset-light wedge, assuming 3 to 8 enterprise customers paying meaningful setup fees plus recurring software and services revenue. Physical-infrastructure versions run much longer and lean heavily on outside financing. This break-even figure is speculative — treat it as a shape, not a promise.
The audience you’d need first. Roughly 200 reachable decision-makers — not customers you’ll go cold-find, but people you can already reach or get warmly introduced to — across U.S. data-center developers, large-load project developers, independent power producers, battery and microgrid firms, and industrial power-asset owners who both feel the bottleneck and can champion a pilot. The founder must already be genuinely connected to this world, or add an estimated 12 to 24 months of reputation-building and conference grind before real sales chances appear.
What moves the needle. For a cold stranger, this is a 1-star business. For a founder who already has deep trust with the exact buyer set, it rises to 2 stars. With both money and the right audience, still 2 stars. And here is the ceiling: no amount of money or audience can push this past 2 stars for this broad variant, because the structural walls — long trust cycles, giant incumbent competition, and moats that live in power access and financing rather than software — do not move.
Who should build this — and who shouldn’t. The person who wins here is a patient infrastructure operator: someone who has spent years around utilities, project development, data-center delivery, microgrids, or industrial energy finance, is comfortable with bespoke deals and heavy diligence, can call developers and power owners directly, and has enough runway to survive 12–24 months of credibility-building before the business becomes legible. The person who’ll be tempted but shouldn’t be is the talented software builder or AI marketer who sees a giant market chart and assumes the hard part is a dashboard or an optimization model — someone who has never sold into power, infrastructure, or industrial procurement and underestimates how much trust, liability, and financing shape every purchase.
Go-to-Market — where the tiny buyer set actually is
This market does not live online; it lives in a small circle of people who award multi-million-dollar power decisions to operators they trust. Three channels fit, and only three.
Direct account-based sales (strong fit). Build a list of 50 to 100 specific buyers — data-center developers, large-load project developers, independent power producers, battery-and-microgrid developers, and industrial landowners with power — and run founder-led outreach with a concrete, site-specific offer: an interconnection-acceleration audit, a behind-the-meter capacity-monetization study, or an inference-siting feasibility package. Cost is roughly $1,000 to $5,000 per serious opportunity in founder time, travel, and engineering support. First 90 days: 30 buyer interviews, 10 tailored proposals, 3 pilot-scoping conversations.
Partnerships and co-marketing (strong fit). Secure 3 to 5 channel partners who already hold the buyer relationships a stranger lacks — industrial-control-system vendors, engineering-and-construction firms, battery integrators, utility consultants, or project-finance boutiques — and package the wedge as an add-on that helps those partners close and accelerate projects. High relationship effort, low media spend. First 90 days: map the partner ecosystem, sign 1 to 2 design partners, build one co-branded offer.
Events, associations, and trade shows (strong fit). Show up where this tiny market physically gathers — data-center, microgrid, utility large-load, and distributed-energy conferences — to compress months of cold outreach into dense relationship cycles. Cost is roughly $10,000 to $50,000 across light attendance, travel, and follow-up over 90 days. The success metric is meetings and pilot conversations, not impressions.
One thing a day: send one sharp outreach email, book one buyer or partner call, publish one niche insight memo, request one warm introduction, or refine one project-specific proposal.
Considered and set aside: SEO/content, paid search, paid social, influencer, own podcast/YouTube, community, guest podcasting, newsletter swaps, affiliates, press/HARO, launch platforms, awards, app stores, listing sites, DMs as a primary motion, local press, freemium, built-in virality, and generic referral programs.
And no marketing rescues this. If the base product has no moat and sells into 6-to-18-month trust cycles against $10-billion incumbents, a clever channel plan just gets you to “no” faster and cheaper.
An illustrative case study
Illustrative scenario — not a real company. Picture “Nadia,” a former interconnection engineer at a regional utility who spent nine years watching data-center developers wait three years for grid connections. Nadia is not a coder; she is a process operator who knows, by first name, a dozen developers stuck in the queue and a handful of battery integrators who could shave a year off energization. She raises $2 million, hires four people, and picks one razor-sharp wedge: a paid “behind-the-meter AI inference siting study” priced at $30,000.
The gamble: Nadia’s first three months are pure outreach — 30 interviews, no product. Five developers say the pain is urgent now. Three share real project data under a confidentiality agreement. One pays for a study before any polished product exists. That single paying pilot is the whole ballgame — proof that a buyer will trust her judgment on a revenue-critical power decision. Had zero paid, the honest move would have been to kill the idea at month two rather than spend a year building a dashboard nobody would buy. Nadia’s edge was never software. It was the phone numbers in her contacts and the years of scar tissue behind them — the exact asset a cold stranger does not have.
The Skeptic’s Case
The strongest reason not to build this is structural, and it is close to airtight. If you stay light enough for founder economics to work — pure software, brokerage, orchestration — you almost certainly lack the moat and the trust to win mission-critical infrastructure business. If you build enough moat to matter — secured power, permits, financing, delivery capability — the founder economics break into tens of millions of dollars and the timelines stretch past most people’s patience. The buyers who matter do not want a clever outsider; they want proven operators who can shorten energization timelines without adding reliability, contract, or financing risk. That collision is the whole trap.
And what we could not verify is telling. Several of the most exciting “distributed” examples — Verse, Armada, Antimatter, and Sunrun’s pilot — rest largely on the companies’ own announcements, and Sunrun itself calls its distributed-compute effort a pilot. We found no strong independent proof that the distributed thesis has produced a breakout startup wedge. The opening may exist; the proof is still thin.
The Thesis-Killer Alternative
Here is the obvious question: isn’t there an easier path that makes this whole business unnecessary? Yes — and it matters. The costly core requirement is obtaining dedicated, reliable, financeable power fast enough to monetize AI demand. A buyer can bypass all of it: lease capacity from existing large clouds, colocate in an existing campus, buy flexible power straight from incumbents, or simply wait for the giants to absorb demand. National Grid Ventures committed $1.75 billion toward accelerating power solutions for U.S. data centers, one of many incumbents muscling into exactly this gap. (nationalgrid.com)
Does the bypass defeat the opportunity? For the infrastructure-owner version — the “we’ll also build AI campuses near power” pitch — yes, the bypass mostly kills it; the giants have that ground. For the enablement version — helping customers jump the queue or monetize idle distributed power faster than incumbents can standardize the workflow — no, not fully. The need is real. But “not fully defeated” is a thin ledge to build a company on, and it does not lift the rating.
Business model, the fastest way to learn the truth, and the risks
The model. The credible early revenue shapes are: development-and-orchestration software plus setup fees sold to developers; a managed “speed-to-power” service with success fees tied to faster energization; or capacity-reservation brokerage between power-asset owners and AI compute buyers. Average first-year customer value runs an estimated $250,000 to $500,000, which is why roughly 6 to 8 customers get you to break-even.
The fastest kill test. Within 45 to 60 days, run 20 to 30 interviews with named target buyers and partners, then try to close 3 paid diagnostic pilots at $10,000 to $50,000 each for one very specific wedge — say, a “behind-the-meter AI inference siting study” or a “large-load queue-bypass power plan.” Kill the idea if fewer than 5 buyers call the problem urgent now, fewer than 3 share real project data under a confidentiality agreement, or zero will pay for a pilot before a full product exists.
The risks, plainly. Capital intensity if you drift toward owning anything; relationship dependence if you touch utilities or large-load procurement; long sales cycles; high trust barriers; and a field where the obvious territory is already occupied. The reasons not to pursue it are the same as the reasons the score is a 1: it is hard to prove, slow to sell, low-moat, and still capital- and relationship-heavy — while the higher-moat versions belong to infrastructure-scale incumbents.
The Full Top 3 Takeaways
1. The market is a five-star fire, and the startup on top of it is a one-star trap. Data-center electricity use nearly tripled from 2014 to 2023 and could reach 9% of all U.S. power by 2030 — the demand is measured, not imagined. But a hot problem is not a good founder bet. The accessible version of this business is hard to prove, slow to sell, and weakly defended, which is exactly why the market’s heat does the founder no favors. The open question that will decide it: does the shortage stay severe long enough — and do interconnection queues stay brutal enough — that buyers finally reward a nimble outsider over waiting for a giant?
2. The moat lives in power, permits, and utility handshakes — none of which a coder can type into existence. The person who wins here is a patient infrastructure operator with a real Rolodex, not a software builder with a slick demo. That is a warning, not a detail: the skill that feels like the job (building the product) is not the skill that closes the deal (earning trust with revenue-critical power decisions). The cliffhanger: will the “distributed” enablement layer ever standardize into a repeatable, sellable product — or will every deal stay so bespoke that no software moat can form and the value keeps flowing to whoever owns the megawatts?
3. The giants already own the obvious ground, so the only remaining wedge is a thin ledge. Crusoe past $10 billion, CoreWeave with $2 billion from NVIDIA, Applied Digital with a $5 billion financing line — the campus land grab is a late-inning game. The enabling layer is earlier, still 2nd inning, category boundaries still unstable. The genuinely unresolved question that will settle everything: does inference — the everyday work of answering AI requests — grow big enough and tolerant enough of being spread across many small sites that a company blending distributed power, modular data centers, and orchestration software becomes a real platform? Or does centralized efficiency keep winning, leaving the distributed dream a permanent niche workaround? Nobody knows yet — and that answer, not the market’s heat, decides whether anyone should build this at all.
Method: Super Fluidity Opportunity Dossier, Fluidity Rating v1. This dossier was produced by an AI analyst system. Market, funding, and competitor facts were source-verified against primary and independent reporting where available and labeled by confidence; figures marked estimate or speculative are modeled, not measured. Illustrative scenarios are labeled projections and do not describe real companies.
Why five ratings? Because the same idea is a different bet for different people. As a side gig — no money, no audience, no special skill, a few hours a week beside a job — the question is simple: is there a realistic path to $2,000–$5,000 a month within a year? As a business, starting with no money and no audience, most ideas are traps — even ideas that became famous companies rate only 2 stars from that starting line. With money, an audience, or both, the same idea can be a real opportunity. So every dossier rates all five, and you can find your own row.
Dossiers rated 0–2 stars for a founder with no money or audience are published free, right here. The rarer 3-, 4-, and 5-star finds go to paid subscribers first — that tier is opening soon, and free subscribers get the first invitation. Join the free list. Or have us investigate your own idea, privately: Request a Dossier ($199).
