But raising capital isn’t necessarily a binary. It is a spectrum. The instruments along it are really just different shapes, different ways of slicing up risk and reward, and the right shape depends on what you are actually building and financing. Funding a science experiment is fundamentally different than funding the 100th copy of a machine you already know how to make and it works, and you shouldn’t finance them with the same instrument. Once you see the full spectrum, structured capital is about picking the shape that fits.

Think financial instrument as a Lego set, there are only two building blocks: a call and a put. Owning equity is a call. You get the right to everything above what the company owes, and if the whole thing goes to zero, you walk away. That walk-away right through a LLC entity is a put you hold. Lending is the exact mirror image. A loan is a risk-free bond minus a put. A lender has effectively sold the founder downside protection, agreeing to take the company assets at a floor price if things go wrong. A floor is a put. A slice of upside is a call. Most deals on earth are just some combination of those two, with the strikes and portions of the reward negotiated to parties’ taste.

It took me a while to understand why “getting a loan is long a put”. A put is the right to sell something at a fixed price, no matter how far its value has actually crashed. Now think about what a borrower gets to do. Say you lend a company $100. If the business thrives, it repays your $100 plus interest and keeps all the upside. You get none of it. But if the business collapses and the assets are only worth $50, the borrower defaults, hands you the wreckage, and walks away. They have effectively sold you the company for $100 when it’s only worth $30. That right is a put. The founder bought it. The lender wrote it. The interest on the loan is the premium the lender charges for standing behind that obligation. This is exactly why riskier borrowers pay higher rates. The put they hold is worth more, so it costs more.

And because with those two building blocks, you can stack calls and puts at different strikes, sizes, and seniorities to manufacture almost any payoff you want. For illustration purposes, a handful of the shapes that come up often.

On the far left sits the flattest shape of all: project finance. This is how most of the actual clean energy build out gets funded. It is non-recourse, secured against a single or a portfolio of bankable assets, and it’s usually cheap because the put you wrote on a contracted, operating asset is expected to be so far out of the money. Senior secured debt keep that capped, flat profile but with recourse to the whole company. Mezzanine debt sits more junior. It takes on a deeper put, charges a higher coupon, sometimes grabs a small equity kicker, and the payoff line begins to bend slightly upward at the top.

Then you reach the middle of the spectrum. This is where I spend a lot of my time. Structured debt with warrants is, in options terms, a bond minus a put plus a tiny call. You get a secured floor that protects the capital, and a warrant that lets you participate if the company takes off. The shape looks like a protected call. This shape exists because a very specific kind of company exists, with assets that are near bankable. They are deeply in money to be priced like a naked call, but not boring enough for the flat project-finance shape on the left. Think of a geothermal developer with validated reservoir but more wells still being drilled. An industrial-heat company with units running beautifully at 1-2 sites, staring down a contracted pipeline 10x larger that it can’t fund off its balance sheet. A modular water-treatment business that works perfectly but can’t finance the next 10 installs. A distributed-battery operator that needs hardware running in thousands of homes before the cash flows truly compound. The combo instrument fits them beautifully because it prices the proven part like debt and the unproven, rapidly scaling-up deployment part, like a small call.

Right around here you meet the convertible, the classic way to raise an unpriced round. The whole trick of a convertible note (or a SAFE) is that you take the money now and do not have to argue about valuation today; it converts into equity at the next priced round, when there’s more information and a lead to set the number. In options terms it’s still a floor plus a call — you’re a creditor until conversion, then you convert into the upside.

Keep going right, preferred equity, and we need to pause, because “preferred” is not one thing. It spans most of the spectrum. PE-style structured preferred often has a redemption right, a contractual obligation for the company to buy the shares back, and has a coupon that accretes the redemption value over time. That’s an engineered floor, often against a business with real cash flows, which is why it leans toward the debt end. Venture-style preferred has only a 1x liquidation preference: a priority claim on whatever exit proceeds exist, junior to every dollar of debt and senior to common. It protects you in a soft landing, where the company sells for less than its price but more than zero and you get your money back before the founders get theirs. It is worth nothing in the outcome that actually dominates venture, when the company goes to 0, 1x of 0 is 0.

At the very end of the line sits common equity: a naked call. All upside with no floor, permanent and fully dilutive. The most expensive shape there is.

AI has bent the entire market toward that right side. Capital is flooding toward companies that serve it. But a naked call is a bad shape a near-bankable company can sell. It gives away for free the floor those companies have already earned by building real assets and signing real contracts. So when a founder asks me what structured capital is, the honest answer is how much of the floor you’ve already earned, and we can go design a shape from there. And for a company that has built real things and building more, the right shape may not live at either edge.

I was staring at a cap table this week when the analogy clicked: the U.S. has finally decided to field a national team in critical minerals.

For two decades the West ran a free-market lineup against a state that plays as one coordinated team: China is the coach, the scout, and the owner/bank all at once, and it doesn’t sweat quarterly earnings while it locks up the next century worth of supply. It’s hard to beat that game when you’re a Series A - a talented high school team — going up against a national squad, even if your technology is Messi-like.

For 15 years the policy works on the cost curve — subsidize the tech, drive down the LCOE, wait for parity. The IRA was that logic at full volume. Recent market signals start to shift to a different direction: policy has shifted from subsidizing technology cost to guaranteeing demand and price. The government just became the long-term offtaker of last resort. That single move drags the bankability question out of the chemistry/commodity question.

1. $110/kg — the price for neodymium-praseodymium the Pentagon agreed to pay MP Materials no matter what the market does. A hard floor set last July and now, per Reuters, the DoD is preparing to copy it across five to ten more minerals. Heavy rare earths, antimony, graphite, tungsten, the shopping list of everything China can choke off with one export license. The mechanism is the Defense Department’s “Open Price” tool, and the ambition is to turn a one-off bailout into a standing regime.

2. The G7 met in Canada stood up its own version - a critical minerals alliance with a price stabilization platform, piloting on lithium and nickel.

3. First Phosphate locked in offtake agreements and LOIs spanning its mine in Quebec through to phosphoric acid for LFP batteries. That is a buyer committing before the plant exists.

4. American Battery Technology won its appeal and got a $57M DOE grant reinstated for its $115M claystone lithium refinery at Tonopah Flats, a project the agency had killed last year, back from the dead.

5. The DoD’s Office of Strategic Capital (what a name!) signed a $500M conditional loan to Phoenix Tailings inside a ~$1B package to build its “Freedom Facility” — domestic rare earth separation and metallization, targeted for 2028.

People ask why domestic rare-earth and lithium refining hasn’t been built. It isn’t that nobody will write an equity check. It’s that every time a Western competitor nears commercial scale, China floods the market, drops spot below marginal cost, and the revenue line your debt was sized against just vaporizes. You cannot project-finance a plant whose output price Beijing controls.

A 10-year purchase agreement at a fixed floor changes the math entirely. You’re no longer betting on commodity prices. You’re betting on the U.S. government honoring a contract (which is not a zero risk). But a government contract is generally bankable. A spot market China can manipulate is not.

Ideally you don’t need the floor as the companies you back have structural and durable cost advantages. Those are the most defensible.

Two names sit at the top of my curiosity list, and both win on feedstock.

• Phoenix Tailings runs a near-zero-waste process pulling rare earth metals out of existing mining waste. So feedstock is already aboveground, already paid for, no a decade long of permitting fight to dig a new hole.

• Cyclic Materials does the recycling-side version: it recovers rare earths from end-of-life EV motors, drones, and robotics. Based on a simple insight, the magnet is a tiny sliver of a scrapped motor by weight but carries nearly all the critical-mineral value. You move roughly 2% of the mass to capture the part that matters.

If your input is waste or scrap, your cost basis is disconnected from global ore markets, which improves bankability. Ideally, the companies stand on commercial merit, and the government still write the anchor offtake or the check.

Back to football. Fielding a national team was the right call IMO. But the players still need their own legs under them. Goal USA, sure. Just make sure we built talented strikers, not a team that needs a penalty kick for every goal.

Storage tech is a stack, sorted by hours of discharge and maybe costs. Every layer has a different job. I’ll argue the case: the last 5% is where the money will be made. Let me explain.

For the bottom layer, daily shifting to move noon solar into the evening and overnight, LFP has basically won the whole board. It took about 95% of utility-scale battery awards in 2024 and 2025. Cells are trading at $55–75/kWh right now, and the EIA is projecting 24 GW of US additions just this year. I still get pitched NMC batteries for grid projects occasionally, which makes zero sense to me. Yes, NMC built the EV industry. Its whole edge is energy density. If you’re trying to cram capacity into an EV chassis where weight and space are binding constraints, great. But a grid battery is literally just a container sitting on a lot discharging over 4 hours. Density buys you nothing. All you get is the massive cobalt bill, half the cycle life, and a fire record (we all remember Moss Landing was NMC).

There is a new entrant at the daily layer: sodium-ion. It competes purely on cost per cycle, using incredibly cheap, abundant materials and boasting 15,000+ cycle claims. The smartest thing CATL did was build their new sodium cell on their existing lithium platform so integrators can swap it in without retooling lines. Makes that 60 GWh sodium order they booked in April look very real. Peak Energy shipped the first US grid-scale sodium systems last year. GM stepping in as a cell partner with them is a great signal.

If you stack solar with 4-8 hours of this stuff, you get close to firm power. A Berkeley team ran 8 years of hourly weather data across 68 sites and found that solar-plus-storage covers over 95% of a data center’s demand at $60–138/MWh. IRENA is putting firm solar-plus-storage at $54–82/MWh in sunny locations. That outright beats new gas, but you take on the long tail of weather risk. I know the prevailing argument in the market is data centers/AI factories are cost insensitive, but over time, the marginal cost of AI might be closer to the marginal cost of electricity.

Now let’s talk about the last 5%: the three-day cloudy stretch; extreme weather events, you can’t just oversize daily batteries to cover that; the capex gets absurd. This is where multi-day storage comes in, and the economics completely invert. Because the asset sits idle most of the year, round-trip efficiency barely matters. Raw cost per kWh of capacity is the only metric you care about. Form Energy’s iron-air battery is intentionally slow and inefficient, targeting $20/kWh with 100 hours of discharge. Google buying that first big project—300 MW / 30 GWh for a Minnesota data center, paired with 1.6 GW of new wind and solar—is huge. It’s the first time anyone has actually bought that last 5% as a standalone product instead of just giving up and throwing in a gas turbine.

I still haven’t figured out the squeezed middle. The 8–24 hour technologies like flow batteries and compressed air. The job is real (bridging a single bad day), but LFP keeps getting cheap enough that developers just buy more hours of lithium, and the 100-hour guys are potentially locking up the multi-day bids. The US players stuck in the middle are bleeding out.

China’s dynamic is intriguing right now. They’ve de-risked nearly every component of this stack, but they aren’t assembling the stack for a reason. Cumulative storage hit 144.7 GW by the end of 2025, up 85% in a single year. They have flow and compressed-air plants at hundred-MW scale operating, not just piloting. The Baochi station in Yunnan, the world’s first grid-forming sodium-ion plant at 200 MW / 400 MWh, went up in 7 months. Seven months! But it’s all grid-side, firming wind and solar farms. Chinese data centers don’t firm on-site because they don’t have to: their grid is strong, interconnection isn’t the bottleneck, and they meet the 80% green-power rule mostly through grid purchases and siting near renewable bases. The national compute-power coordination policy push trying to turn data centers into flexible, dispatchable loads, is still mostly aspirational. Nobody has figured out how to map compute tasks to power load precisely enough to dispatch against it yet. So we have this fun split: China de-risks the hardware, and the US actually pieces the complex systems together under duress, purely because our grid failed the load.

Can stacking storage tech make solar firm? Mostly, and cheaply. The daily layer gets you 95% of the way there. For the last 5%, there are really only three sellers: gas (tracking 101 GW of behind-the-meter announcements right now, and Bloom’s stock went up by 1000% YoY, a 10x that everyday people could actually access), 100-hour storage (just the one Google order so far), or the grid itself.

The grid is the ultimate top layer of the stack, infinite duration, deepest firmness. But it’s honestly a deeply underpriced insurance product. Because it’s egalitarian by design, utilities recover costs at a fixed, regulated margin spread across everyone. A self-supplied load draws on it exactly when everyone else does. That’s exactly what those ugly cost-allocation fights in Texas, Tennessee, and Oregon are about: regulators trying to reprice that standby product mid-game. Imagine if Kalshi could just change your contract price with two seconds left in Game 5 of Knicks–Spurs you’d scream it’s unfair. But the grid has no other option, the costs are real, and somebody has to recover them.

If the stack can flex or even export, we stop buying insurance from the grid and start selling it. Figuring out the last 5% is so exciting, it's the only slice of the stack nobody owns yet, and there will be real winners!

Other things I’m keeping an eye on this week:

• China’s grid added 18 GW / 65 GWh of storage in a single month late last year. That’s a quarter of the entire global 2025 deployment in thirty days. Insane scale.

• GM’s Ultium JV is spinning up LFP production for grid storage in July, with sodium prototype cells due by year-end. Using idled EV lines to absorb storage demand is smart.

• Inlyte Energy is commissioning a 600 kWh iron-sodium pilot inside a Tier IV Swiss data center late this year. Need to check their data when it goes live.

And the public market is opening its doors much earlier than it used to. Fervo priced its IPO on May 13 around $7.6B and opened near $10B, the largest clean-energy listing on record. Fervo booked $138K of revenue last year. Two weeks later, the Italian fast-reactor developer newcleo confirmed a Nasdaq SPAC merger at a $2.4B valuation; underground-reactor startup Deep Fission is running a Nasdaq roadshow targeting up to $1.7B. Newcleo’s first commercial reactor isn’t penciled in until 2032. The market is now willing to fund things that are pre-revenue, sometimes pre-prototype.

When every MW clears at a price that assumes the boom, and every M&A process opens at a billion, if I’m a mid-market infra or growth fund writing $30–200M checks, I’m not winning that auction. I’m bidding against utilities with a regulated cost of capital, credit shops with a lower one, and now a public market that’ll fund pre-revenue. So the real question for the month: when demand is no longer a secret, where is anything still underpriced?

Been exploring 3 things. First: own the bottleneck, not the MWs. The scarce thing was never power, it’s *deliverable* power. A wave of companies is attacking the “where can I actually get electrons” problem from every side: EPCs, construction labor, the grid equipment supply chain. Texture raised $12.5M to instrument the rural co-op grid (42 million Americans on systems built for one-way power flow). DG Matrix raised $60M to build solid-state transformers. Heron Power launched partnership with LG Energy Vertech. You’ll build your own list.

Two: Manufacture the credit, own the spread. The deal I found most interesting this month wasn’t exactly a deal. Eos and Cerberus launched Frontier Power USA for long duration zinc-based storage, Cerberus anchoring $100M. But the equity isn’t the story. The story is the ~$1.5B, 15-year, non-cancellable technology-performance insurance framework from Ariel Green , backed by A+/AA- Lloyd’s paper, that lets the project debt price at investment grade, turning squishy technology risk into rateable credit risk. Love the structure.

And that’s not the only way to manufacture credit. On June 2, Voltus and Google signed a first-of-its-kind “Bring Your Own Capacity” deal — 100 MW of aggregated distributed resources in PJM, three years, with Google funding the VPP so the capacity shows up as clean, dispatchable supply. Same idea as the Lloyd’s wrap: somebody steps in to swallow the messy, un-bankable risk so the financier gets investment-grade paper. Whoever owns the credit-enhancement template can earn the spread of an asset class.

Three: Buy the brownfield, fix the back office, repower. A decade of DG got built for the tax-equity flip, not for 30 years of operation. This is merely a hypothesis: the returns are dragged because nobody’s minding the store. Raptor Maps clocks average site underperformance at 5.8%+ and that’s just the production. The drag also sits at back office burdens for DG assets: unbilled community-solar credits, subscriber churn, RECs left on the table, claims never filed. So roll up 20+ sub-5MW sites, automate the back office onto one platform, enhance cashflows. Then repower where it pencils, an inverter-and-module swap runs 40–70% of a new build, lifts yield 10–30%, and keeps the interconnection. With new queue positions 4-7 years out, what you’re really buying is a live grid connection?

Will report back on what I find along those threads. It does feel like a market that buys power at any price isn’t the same as a market that’s easy to make money in, but I do see the middle as the one place left to be early on something other than demand.

4 years ago I spent a lot of time researching battery material recycling. This was right as the supply chain started bifurcating along geopolitical lines, and there was an obvious sustainability story. But there was also a national security story that felt more durable: if China controls the midstream of refining and cathode production, then recovering critical minerals from spent batteries is industrial sovereignty.

I watched the category closely for a long time. And I kept running into the same problem. Ascend Elements (filed Chapter 11), Li-Cycle (sold for $40M), Redwood (pivoted to secondary-life as storage), these were genuinely hard companies to underwrite. Enormous capex, technical complexity, beautiful chemistry yet difficult to validate at thousand-tons scale. I assumed that this was a scaling problem. Get the volume up, ride the cost curve down, and the math would eventually close.

Then a physicist friend changed how I think about the entire space.

In a closed system, entropy always increases. Energy is conserved, but its grade only degrades from useful concentrated forms toward diffuse, less useful ones. A hot cup of coffee cools. A charged battery drains. Heat flows from hot to cold. A finished cathode is a very low low-entropy object: specific elements in specific ratios in a specific crystal structure. When you shred a battery into mixed black mass, you move up on the entropy gradient, and the Second Law says you have to pay for the trip back down in high-grade energy. The more dilute and mixed-up your feedstock is, the steeper that climb gets. Scale doesn’t save you. The toll is the toll. To justify the toll, the end product needs to be highly valuable to pay for the bill.

Once you have the lens, you see it everywhere. DAC is the same trade. CO₂ in air at 420 parts per million is so diluted, which is why the thermodynamic floor sits near half a gigajoule per ton and real systems run 5-10x that. The energy bill to pay for entropy decrease is extreme. The DAC projects that might eventually pencil are probably the ones colocated with energy nobody else wants, stranded geothermal heat or curtailed renewables.

But I still love recycling, maybe not the ones fighting against the laws of physics. I found another form of recycling!

When I was at school, I won a startup pitch competition with an embarrassingly simple idea. Office buildings have parking garages in prime locations that sit empty every evening and weekend. So rent that idle capacity to the public during the downtime. I pitched to recycle a physical asset’s slack, when an expensive, well-located thing was doing nothing.

Years later, I keep seeing the same formula at much larger scale and executed by talented teams. Voltus aggregates spare flexibility in commercial and industrial loads and turns it into market-accredited capacity, now managing 8 GW+ and pitching data centers a “BYOC” path to faster interconnection. GridCare uses AI to find the roughly 2/3 of grid capacity that sits unused under normal conditions and hands it to data centers waiting years to connect. LineVision does dynamic line rating, recycling the headroom that already exists on transmission lines once you actually measure what they can carry instead of assuming a conservative static limit. They all recycle slack in something that already exists.

So is there a Second Law for this kind of recycling too? I think there is, and maybe called the law of motivation - businesses are motivated by profit. These models all require multiple independent parties to agree to share an asset they would otherwise own and hoard. The single most reliable solution for that kind of coordination problem is profit.

For years, this category struggled for exactly that reason. Selling a flexible load agreement to a C&I customer when the only counterparty is a utility or a regulator is a dead end. Neither is profit-driven, and the motivation to share was thin. Genuinely good products died for the law of motivation.

What changed is that data centers showed up willing to pay enormous sums for fast access to power. That single new buyer rewrites the coordination problem. Now there is real money to split, so the C&I customer has a reason to commit its idle capacity, the utility has pressure to cooperate, and the aggregator finally has margin to operate inside. The slack capacity was always there. What was missing was a party hungry enough to fund the alignment.

Two conservation laws: One says you cannot recover concentrated value from a diffuse system without paying in energy. The other says you cannot (sufficiently) motivate multiple independent counterparties without paying in profit.

Food is costing more, and investing in climate-friendly food systems has cost probably even more, because the unit economics don’t compete. Beyond Meat couldn’t make a burger cheaper than Tyson. Bowery couldn’t grow lettuce cheaper than California. AppHarvest burned $792 million and still couldn’t hit commodity pricing on tomatoes. These companies needed consumers to pay a green premium, and consumers wouldn’t.

But the structural forces driving food prices up have gotten worse. Fertilizer is up 15–43% yoy. 61% of the US is in some kind of drought. The Iran conflict is keeping fuel costs high. Extreme weather events, energy costs, and geopolitical fragmentation are saying cheap food aren’t coming back. The demand for sustainable food production should be real and growing. Maybe the question is how to “productize” it, turn new food production into something financeable.

Start with the 3 questions project finance always asks: Is the unit repeatable? Is the revenue contractable? Is the capex predictable per unit of output? Companies say yes to all three look less like expensive alternatives to commodity ag and more like manufacturing platforms with contracted revenue.

That points us toward biomanufacturing. Precision fermentation and engineered biology have been around for years. The good news is ML is compressing the R&D cycle, strain selection and bioreactor optimization that used to take months now takes weeks. The bad news is the same old, scaling production from 10 liters to 10,000 liters is super high risk, and raising capital to attempt it is even harder. But without scale, you can’t hit your cost curve. Every new biomanufacturing company faces the same brutal question: how do you bridge from pilot to commercial ($100M) at the lowest cost of capital?

So why not build shared biomanufacturing infrastructure? The iFAB consortium in Illinois is an interesting experiment. It is an $81 million effort backed by the Department of Commerce and the state of Illinois, brokering access to ADM and Primient’s existing large scale fermentation capacity for food-grade startups. Their thesis is straightforward: expecting every startup to build its own fermentation plant is untenable. BioMADE, a DoD-backed consortium, is building pilot-scale facilities in California and Iowa with the same logic.

Pharma is the obvious precedent. CDMOs, contract development and manufacturing organizations, are a $25 billion-plus industry. Lonza, Samsung Biologics, Fujifilm Diosynth: these are shared biomanufacturing at massive scale. Samsung alone runs 604,000 liters of capacity serving 17 of the top 20 pharma companies. It works because pharma products are high-value enough to support premium tolling fees, and the regulatory framework actually standardizes facility design in ways that help multi-tenancy.

Can the same model work for food ingredients? The data center analogy is tempting, but shared compute works because a byte is a byte. Biology isn’t that clean. Every organism has different fermentation conditions, contamination profiles, downstream processing. Switching tenants means cleaning, reconfiguring, revalidating. That’s real cost and real downtime. But maybe you don’t need full fungibility. You just need enough standardization at the most expensive step. Fermentation vessels are the capex heavy part, and a 10,000L bioreactor doesn’t care whether it’s growing yeast for protein or fungi for collagen. Leave the downstream processing to tenant-specific. From an investor’s perspective, the most compelling thing about shared infrastructure is shared FOAK risk. For a Series A company trying to raise, the easiest way to answer the “can you actually scale 100x?” question is: we rented time at an existing facility, ran full production cycles for 6 months, and we’re ready to build our own. That de-risks the tech by magnitude. You’re not asking an investor to bet on growing an organism at a scale that’s never been tested. One close example is Michroma ferments fungi to produce natural food colorants for food manufacturers, high-value, B2B, stable demand as the FDA phases out synthetic dyes. But building their own fermentation plant costs $200 million. So they partnered with CJ CheilJedang, one of the world’s largest fermentation companies (you know them if you eat kimchi), to produce at commercial scale.

What’s smart about iFAB is also that it isn’t building greenfield, but leveraging existing facilities at companies that have run fermentation for decades. I honestly do not know whether food ingredient margins can support the coordination overhead and whether the startup tenant credits can support financing for such shared facilities yet.

This is all half-baked thinking, a detour from this week’s Fervo massive headline and more data center talks, maybe a few years early for where the biomanufacturing industry actually is. But S2G just closed a fresh billion dollars, and if anyone wants to build shared infrastructure for the next generation of food systems, you know where to find them :)

So how do we solve this? The simple and conventional answer is that hyperscalers pay a premium to utilities, and utilities credit it back to affected households. Maybe call it AI energy credits. It’s clean, it’s politically intuitive, and it’s pure redistribution — the kwh don’t move, only the dollars. Physically nothing changes. The grid still cracks when a training run hits at 6pm on a 100-degree day.

Or, creatively, hyperscalers buy four-hours of battery for every household affected. Think about what that does. The homeowner gets a battery they didn’t (directly) pay for. They get backup power during the next brownout. Their bill goes down because the battery clips their on-peak draw. The grid gets dispatchable capacity sitting at the edge of the distribution system, exactly where the new load is landing. And the hyperscaler instead of writing a guilt check to subsidize someone else’s gas bill gets accredited capacity that PJM can count toward reliability. Same dollar, 4x of value.

But hyperscaler CFOs will struggle to explain a multi-billion-dollar residential-battery line item on the next earnings call, and most homeowners aren’t going to take a free Powerwall from Amazon. The structure needs an intermediary — somebody to procure the assets, hold them on a balance sheet, operate the fleet, and stand between Big Tech and the homeowner. Who’s actually built for this job?

Ouachita Electric Cooperative in southern Arkansas has 13,000 members and a service territory that overlaps with some of the poorest counties in America. Since 2016, it’s been paying upfront for heat pumps, insulation, and weatherization in member homes, recovering costs through a small tariff on the monthly bill that’s guaranteed to be less than the energy savings. No credit check. No lien. No debt. The customer saves money from Day 1. 700 homes later, Ouachita has cut summer peak demand by 30%, reduced average household energy use by 18%, and contributed to a 4.5% rate decrease for ALL members, including those who didn’t participate. The program’s loan loss reserve, backstopped by the Arkansas Energy Office, has never been touched.

Lightshift Energy and the Blue Ridge Power Agency announced 25 MW of distribution-connected batteries across three Virginia co-ops three weeks ago, expected to save those utilities $100 million over project life. Lightshift owns the batteries, operates them, and shares peak-shaving savings through a 20-year service agreement.

There are roughly 2,800 electric cooperatives and municipal utilities serve 24% of U.S. electricity customers, manage critical last-mile infrastructure across 48 states, and have essentially zero battery storage and minimal home electrification programs. They can sign 20-year contracts. And they have a level of customer trust that an investor-owned utility going through a rate case will simply never have. Some of the co-ops are already the counterparty to the hyperscalers. In Virginia (aka the center of AI DC universe), Northern Virginia Electric Cooperative — NOVEC — has 180,000 members. Data centers now account for more than 65% of NOVEC’s energy sales, projected to reach 95% by 2032. Rappahannock Electric Cooperative, the largest in Virginia, expects 17 GW of data center demand by 2040, up from roughly 0 in 2023. The contractual relationship between Amazon and the homeowner across the street already runs through these co-ops. Nobody else in the country has both ends of the wire on their balance sheet.

The capital stack works because of an often overlooked provision in the IRA: Section 6417 — elective pay — allows tax-exempt entities like co-ops and municipal utilities to claim the 30% ITC as a direct cash payment from the Treasury. Before this, co-ops couldn’t access the ITC at all because they have no federal tax liability. That one provision unlocked the entire market. The counterparty is a municipal utility or cooperative with a regulated revenue stream and essential-service billing priority. At 5 MW per installation, a $200-300M fund targeting 40-60 systems could generate 8-12% levered returns with investment-grade-equivalent credit risk. The return could get better if hyperscalers bid higher for capacity.

Now extend it one step. The co-op puts a 4-hour battery in the member’s home. Instead of recovering the cost through the member’s bill, it recovers it through a 10- or 15-year capacity PPA with the hyperscaler down the road. The co-op operates the fleet; the co-op bonds against the capacity contracts. The bonds finance the battery deployment. The hyperscaler books a regulated-utility capacity contract on its balance sheet — much easier to explain to investors. They also get accredited capacity that PJM will count; the member gets backup power and lower peak charges. The unit economics are dramatically better than what Sunrun runs on a single-family basis. The co-op’s cost of capital is closer to 5% than 10%. The contracts aggregate at scale. There’s no customer acquisition cost because these are existing members. NOVEC and REC could each deploy 100,000 home batteries inside their existing service territory and not come close to saturation.

The boring solution to the sexiest infrastructure problem of the decade might just be small batteries on old co-op meters.

P.S. Real thanks to everyone who's been reading and writing back. Half-baked Friday-morning theses/research have turned into actual conversations with old colleagues, new friends, and people across the industry I'd never have met otherwise. It really keeps me going. :)

Start with siting. Data centers have always been sited along fiber backbone on powered land. What changed is scale. A 10 MW cloud DC plugs into existing grid infra without anyone blinking. A 500 MW AI training campus overwhelms what local transmission was built to serve, even at well-connected sites. That’s why PJM and ERCOT interconnection queues stretch 5–7 years and hyperscalers are signing 20-year nuclear PPAs or building gas turbines behind the meter. Memphis xAI is the canonical version: hundreds of megawatts of on-site gas because waiting wasn’t an option. The catch is that on-site generation solves the queue problem and creates a different one — a 24/7 firm load gives you nothing to flex, which undercuts everything in the next layer.

Next is operations, and this is where bidirectional sensing matters. Today, the data center tells the grid what it needs and the grid serves it. One direction. Bidirectional means the grid pushes real-time signals — price, congestion, frequency, carbon intensity — and the data center pushes back its state and how much it can flex right now. Closer to a smart thermostat than a static breaker.

Two physical realities make this hard. AI training has a weird load shape: GPU clusters alternate every few seconds between compute phases (full draw) and communication phases (reduced draw). At 100 MW scale, that’s tens of MWs swinging on a sub-minute timescale. NERC has already flagged voltage excursions traced back to training DCs. The hardware fix is short-duration batteries sized not for grid arbitrage but as power conditioners — smoothing spikes before they hit the transmission line. These cycle thousands of times a year at durations measured in seconds (!!!), not the 4-hour lithium discharge profile most U.S. BESS projects gets built around. Different product, different economics.

The compute side has to bend too: schedulers that pause, migrate, or delay jobs based on grid signals, not at the rack level but at the AI cluster level, which is much harder. Google has been doing a version of this through their carbon-intelligent computing program since 2020, shifting non-urgent tasks toward cleaner, lower-demand windows. Tyler Norris — now Google’s Head of Market Innovation — quantified the upside from his earlier Duke research: if data centers curtail for a quarter of 1% of their annual uptime, the existing U.S. grid can absorb roughly 100 GW of new load, covering most of the AI demand forecast through 2030.

Emerald AI is a Series A company directly building on this layer. Their Conductor platform orchestrates AI workloads against grid signals in real time. Phoenix demo with Oracle: 25% power reduction over three hours without breaking workload SLAs. Their chief scientist’s peer-reviewed paper puts the flexibility envelope across AI workloads at 18–55% depending on job type. They closed $25M in late March led by Energy Impact Partners with Nvidia, and have a 96 MW commercial flexible AI factory in development at Digital Realty’s Aurora site in Manassas with EPRI, Dominion, and PJM. If the grid-DC orchestration play becomes the mainstream, they want to be the operating system where everyone plugs in.

The third layer is the market, and I love this one the most. The pitch is that AI APIs should be priced like electricity, cheaper when the grid is clean and abundant, more expensive when it’s stressed. Time-of-use pricing for tokens. Sounds speculative until you notice it just shipped. On April 2, Google Cloud introduced two inference tiers for Gemini: Flex and Priority. Flex requests can queue for up to 15 minutes and cost up to 50% less. Priority gets guaranteed throughput at full price. I have seen this before, it is called Uber Surge Pricing! Google frames this as cost-and-reliability tradeoffs for AI developers, not grid pricing pass-through. But the structure creates the power-to-token surface through which grid signals could eventually flow. Norris did join Google after all.

Regulatory framing is moving the same direction. SPP has filed a High-Density Interruptible Load tariff proposal. Pennsylvania’s PUC is advancing flexible tariff options for large industrial loads. None of this reshapes the market alone, but the direction is regulators are beginning to frame large loads as participants, not simply demand to be served.

Constellation’s CEO said this at CERAWeek: we don’t have a supply problem, we have a peak problem. If Norris is right, the next 50 GW of U.S. AI load gets accommodated not by more generation but by making compute behave for a few hundred hours a year. I feel that is as much a project finance question as a technical question. As a financier, we always love negotiating risk and reward allocation!

The real bottleneck might be the shortage of people who understand both grid and compute. Then again — in 2026, what can you not learn?

The thesis is really four nested bets:

• Entry price bet - you can buy a distressed installer at a price that leaves room for PE-grade returns after integration pain.

• Cost stack bet — there’s enough AI-addressable cost to move the unit economics materially.

• Technical bet — AI actually delivers that reduction in practice (not just in excel models).

• Demand/upsell bet — lower cost wins volume on the front end, and the installed base opts into dispatch on the back end.

Bet 1: The entry price is below installed-base value. Let’s start with the bodies. Freedom Forever filed Chapter 11 last week — $500 million in debt, 50,000 creditors, the second-largest residential solar installer in the country. A year ago Sunnova went through the same meat grinder: 3+ GW of installed capacity sold to Solaris Assets via a DIP credit bid plus $25 million in cash. SunPower’s operating platform went to Complete Solaria for $45 million. The market for residential solar companies right now is a liquidation sale.

The consensus read: the economics are broken. Wood Mackenzie’s 2026 outlook has residential CAC spiking 40% to $0.84/W. Up to a third of customer payments go to interest at the current rate environment. 6+ quarters of volume declines as consumer-owned tax incentives expire. We are not even talking about tariffs yet. A pure-play installer is an ugly bet and the market is pricing accordingly.

But what if the market is pricing the wrong thing? If I were a YC founder, I would not position these companies as residential solar providers. They’re custodians of hundreds of thousands of rooftop relationships. 20-25 year contracts, on houses with existing interconnection, net metering, and (increasingly) batteries. Sunnova customers changed hands at ~$236 each — the price of a grid-connected, permitted, energized DER with a known load profile and an existing billing relationship, for $236! That feels cheap. A 20-year PPA customer with a battery-ready home has contracted cash flows plus optionality on retrofit, grid services, EV charging, and re-contracting at term. At a 10% discount rate and $40/month net contribution, that’s roughly $4000 NPV per customer — before any VPP layer. Sunnova’s acquirer paid ~6 cents on that dollar.

Bet 2: AI materially reduces go-forward installer soft costs. A typical US residential system runs $3.10-$3.50/W installed (LBNL Tracking the Sun). Soft costs are 40-50% of total — and the biggest line item is customer acquisition at $0.60-$0.84/W, dwarfing actual install labor at $0.15-$0.25/W. The rest are design, permitting, interconnection paperwork, scheduling, all seem AI-able with developed SOP + human in the loop to address edge cases. OpenSolar claims $0.43/W removable from sales and acquisition costs through AI-driven workflows. SolarAPP+ is automating permit review in 160+ communities. Aurora’s AI design tools compress plan sets from hours to minutes. Stack these AI tools together and you pull $0.40-$0.60/W or more out of soft costs, enough to turn a money-losing installer into a cash-flow-neutral one.

Bet 3: The installed base can be turned into a VPP. Home batteries are coming. Sunrun reported 4 GWh of networked storage across 18 VPP programs in 2025, dispatching 425 MW at peak, with 71% battery attach on new installs. Tesla + Sunrun coordinated a 539 MW dispatch to CAISO last July — a mid-size gas plant’s worth of capacity, independently verified by Brattle. Base Power raised $1+ billion to build this model from scratch in Texas.

Three sub-questions hiding in this bet. Retrofit economics: batteries cost $5K-$8K per home to retrofit. What’s the payback against VPP revenue and bill-savings share? Opt-in and dispatchability: You can own the rooftop relationship, but the homeowner has to agree to battery installation and VPP dispatch. how many existing solar-only customers consent to battery install and hand over remote dispatch control to the operator? Inverter and battery vendors vary on how reliably this can be controlled remotely across a mixed fleet (Tesla, Enphase, Franklin, SolarEdge), and legacy Sunnova/SunPower installations weren’t all spec’d with dispatch-grade telemetry. I can’t find published data on the real opt-in number. It could be 20% or 60%, and it changes everything. Wholesale participation: only ~10% of residential VPP capacity currently participates in wholesale markets; revenue per dispatched kWh varies wildly by ISO (CAISO has DSGS, ERCOT has ancillary services, PJM is still writing rules!!).

ISO aggregation rules are still being written in half the country. The thesis doesn’t work so well if aggregation markets stay thin or if tax policy keeps whipsawing. However, we can always start with CAISO and ERCOT where VPP markets are liquid today.

What I would love to do to fully test this thesis by validating those assumptions (1) you can buy below ~$300/customer in CAISO or ERCOT territory, (2) the AI soft-cost stack delivers $0.30+/W of realized compression in practice, and (3) retrofit opt-in among existing customers lands above ~30-40%. Three testable numbers. If they hold, the equity case might be real. Time to call my PE friends!

Converting MWh into token is the most profitable energy business that everyone wants to be in right now. Nearly 2,300 GW of generation and storage capacity are sitting in U.S. interconnection queues, more than the country’s entire installed power base?! Building a 100 MW data center the traditional way takes 3–5 years. The AI buildout doesn’t have 3-5 years. Two different theories are emerging to solve the same puzzle from the opposite directions, and each implies a different view on where value sits in the AI x energy stack.

**Model 1: Bring Distributed Capacity to the Load.** Hyperscalers are starting to contract bilaterally with distributed BESS projects via VPPs or capacity agreements, sidestepping the utility as commercial intermediary. A data center stuck in PJM’s interconnection queue contracts with a portfolio of distributed 4-hour batteries to demonstrate load flexibility and shave peak demand, (hopefully) earning a faster grid connection.

Aligned Data Centers signed with Calibrant Energy for 31 MW / 62 MWh of onsite BESS at a Pacific Northwest campus, sized to accelerate interconnection “years earlier than traditional utility upgrades.” Jefferies estimates hyperscalers represent a 20 GW BESS opportunity through 2035. Wood Mackenzie reports data center demand drove a 33% jump in VPP deployments in 2025, with PJM and ERCOT leading adoption.

The structural frictions are real. You need firmness between the BESS node and the load node. Basis risk (power price difference) between nodes means a battery portfolio in West Texas doesn’t necessarily deliver firm capacity to a data center in Dallas. PJM’s ELCC framework assigns roughly 50% capacity credit to 4-hour storage; you need twice the nameplate to match a gas peaker. ERCOT has no capacity market at all, compensating storage purely through energy and ancillary revenues. CAISO’s resource adequacy rules are tighter but the market is saturated with 14 GW of operational batteries.

Then there’s FERC Order 2222, the 2020 rule meant to open wholesale markets to aggregated DERs. Six years later, implementation is glacial. NYISO targets full compliance by end of 2026; PJM and MISO filings are still under review. Retail tariff demand charges and standby rates add another friction, penalizing the exact load flexibility these contracts require.

This model seems to work, but it’s a sophisticated financial and regulatory arbitrage. The winners will structure bilateral capacity products across ISO seams and manage basis risk through a patchwork of rules. I bet some capable financiers and energy trading desks have figured this out with AI’s assistance.

**Model 2: Distribute the Compute Load to the Capacity.** SPAN announced XFRA yesterday at Latitude Media’s Transition-AI conference: a distributed edge compute system that embeds inference GPU nodes directly into homes. The premise, per SPAN CEO Arch Rao, is that the U.S. distribution grid operates at only 40–45% utilization on average, leaving substantial headroom. Each XFRA Node pairs with SPAN’s smart panel and a whole-home battery; an orchestration layer routes AI workloads across nodes based on available energy and latency. This is fascinating. Can you productize edge compute systems to be DTC shippable?

One hundred Pulte homes (third-largest U.S. homebuilder), 1,600 direct liquid-cooled inference GPUs, 1.25 MW of aggregate compute, deployable in six months. SPAN claims a cost of $3 million per MW versus $15 million for a traditional data center, though that comparison likely favors SPAN: the $15M figure covers a fully built hyperscale facility, while the $3M figure’s treatment of smart panel installation, liquid cooling infrastructure (a non-trivial engineering problem in residential settings, where heat rejection and plumbing maintenance aren’t solved at scale), and whole-home battery costs isn’t clear. Homeowners pay $150/month covering energy and internet; SPAN sells compute (inference-only) to hyperscalers and AI companies. The model sounds like third-party-owned residential solar and batteries, though deploying GPU nodes with coolant loops are a different animal than rooftop panels.

SPAN has been working on this theory for a while. They closed a $163M Series C as of February. PG&E’s SAVE VPP pilot last summer, where SPAN coordinated 400 smart panels and 1,500 Sunrun batteries for localized grid relief.

XFRA is inference-only; you can’t train frontier models across thousands of residential nodes with variable power and latency. Unit economics need new construction, which caps near-term TAM at tens of thousands of homes annually rather than millions. The orchestration challenge of delivering consistent latency across a residential network subject to homeowner behavior and local outages is a hard software problem. Hyperscaler inference SLAs are typically 99.9–99.99%; how close a residential network gets to that is the central question.

There’s also a competitive set. Purpose-built edge micro-data centers from Vapor IO, EdgeConnex, and Compass Edge already offer low-latency inference closer to end users, with commercial-grade power and cooling. The question isn’t just “centralized vs. distributed.” It’s whether inference will fragment to residential nodes further or to commercial edge colo, and hyperscalers already have bilateral deals with the latter.

---

## Where Capital Deploys

These models aren’t competing head-to-head. They solve different parts of the problem, and they imply different capital allocation strategies.

**”Bring capacity to load”** is an institutional play: BESS developers, energy trading platforms, and infrastructure credit funds that can originate bilateral capacity agreements and manage basis risk across ISO seams. The risk is regulatory fragmentation. This is fundamentally a financial engineering problem.

**”Bring load to capacity”** is potentially venture-backable. SPAN’s XFRA is a bet that AI inference will commoditize fast enough that enterprise reliability requirements relax before residential orchestration matures, and that the projected $5T+ in data center capex is structurally exposed to a low-cost distributed alternative. That’s a bold thesis. If it’s right, it creates a new asset class: distributed compute financed like residential solar, operated like a VPP, monetized through inference revenue.

The portfolio question is where along the centralization-distribution spectrum each workload class lands. Training stays centralized. Inference is up for grabs, and the capital structures will follow whoever proves they can deliver reliable compute at the lowest cost per token.

---

## Sources

- Latitude Media, “Span to launch mini AI data centers for distributed edge compute,” April 13, 2026

- Latitude Media, “Span is raising a $176 million Series C,” February 2, 2026

- Latitude Media, “Data centers are beginning to embrace batteries for onsite power,” November 3, 2025

- Latitude Media, “PG&E is testing ‘precision grid surgery’ with this summer’s VPP pilot,” September 19, 2025

- Utility Dive, “In 2026, virtual power plants must scale or risk being left behind,” January 27, 2026

- Jefferies, hyperscaler BESS opportunity estimate (November 2025)

- Wood Mackenzie, VPP market data and data center demand growth (September 2025)

- Advanced Energy United, “What ELCC is Telling Us About PJM’s Capacity Crunch,” March 2026

- Ascend Analytics, “Storage’s Moment: Capacity Markets in Transition,” October 2025

- S&P Global, ERCOT surpasses CAISO in battery storage capacity (September 2025)

- Lawrence Berkeley National Lab, “Queued Up” interconnection data (end-2024)

- FERC Order 2222 Explainer, ferc.gov

- PG&E SAVE Virtual Power Plant Program launch (March 2025)

- Aligned Data Centers / Calibrant Energy BESS deal (October 2025)

- Reuters Breakingviews, “AI dreams crash into stark $7 trln reality,” April 2026

- Sightline Climate, data center delay forecast (2026)

---

*Friday Morning Labs covers climate infrastructure, energy transition, and the capital markets connecting them.*

GridCARE’s partnership with National Grid in New York aims to cut interconnection timelines from years to 6-12 months by modeling spare capacity across the grid in real time. Emerald AI, backed by Nvidia, Energy Impact Partners, Eaton, GE Vernova, and the CIA’s venture arm IQT, closed a $25M round (total $68M in 16 months) to build “grid-friendly AI factories” — software that flexes data center load at peak times so facilities can connect faster without costly transmission upgrades. Their thesis: up to 100 GW of latent capacity sits on the existing US grid if loads are managed intelligently. Soma Energy, founded by ex-AWS energy leaders, emerged from stealth with $7M to do something similar — optimize real-time dispatch between power plants and data centers to unlock spare megawatts.

The common thread: the interconnection queue is 2,600 GW deep. Building new generation takes 5-7 years. But software that identifies and dispatches existing spare capacity can deliver power in months. For investors, this is a classic “picks and shovels” category — platform businesses with recurring revenue, utility partnerships, and capital-light models. The risk is execution: utilities are conservative buyers, and proving grid reliability under flexible load management requires years of operating data. But the demand signal is real, and the first movers are attracting serious capital.

Overheard in the Grid Queue

*This is why three startups raised a combined $100M+ this month to fix grid interconnection with software instead of paperwork.*

## Capital Raises & Deals

**Valar Atomics — $450M Series B ($2B Valuation)**
Palmer Luckey-backed nuclear startup raised $340M equity plus $110M debt for high-temperature gas-cooled reactors targeting data center and industrial baseload. Valar achieved nuclear criticality five months ago and is targeting commercial operations by July 2026 at a DOE-partnered Utah site. The debt component is notable — lenders underwriting nuclear development risk, not just equity. If they hit timeline, it resets the market's view of SMR bankability. Worth watching beyond the reactor itself: high-temp gas-cooled designs run at 750–950°C, which demands specialized heat exchangers, advanced materials (SiC composites, nickel superalloys), and thermal storage for load-following — an entire supply chain layer that barely exists as a standalone category today. As SMRs proliferate across different coolant architectures, “thermal management suppliers” could become the next opportunity, cutting across nuclear, data center cooling, and industrial decarb simultaneously.

**EnerVenue — $300M Series B Extension**
Fremont-based nickel-hydrogen battery maker closed $300M led by Full Vision Capital, bringing on new CEO Henning Rath. The technology targets 4-12 hour discharge with 30,000-cycle durability — infrastructure-grade storage with dramatically lower degradation than lithium-ion. New 250 MWh production line starts construction later this year.

**Bloom Energy / AEP — $2.65B Fuel Cell Deployment**
Up to 1 GW of solid oxide fuel cells for data center projects. Combined with Bloom's $5B Brookfield agreement, nearly $8B in total pipeline. Fuel cells are deployable in months, not years — the speed premium is real for data centers stuck in interconnection queues.

---

## Market Moves

**DOE's $1.9B SPARK Transmission Program — Concept Papers Submitted**
Full applications due May 20. SPARK targets reconductoring — replacing existing transmission lines with higher-capacity conductors on the same rights-of-way. Faster and cheaper than new corridors. The program directly funds the supply chain for companies like TS Conductor and CTC Global. Watch for award announcements soon!

**Long-Duration Storage Hits Inflection Point**
Wood Mackenzie reports LDES installations reached 15 GWh in 2025, up nearly 50% YoY, though still only 6% of total storage. The shift: data centers are becoming the anchor customer class. Form Energy's 12 GWh iron-air deal with Crusoe (deliveries starting 2027), plus a separate Google data center battery in Minnesota, represent the largest committed LDES pipeline ever. Eos Energy says data centers are now nearly 25% of its project pipeline, up from negligible a year ago.

---

## Across the Capital Stack

**Venture: Grid Intelligence May Finally See ARR**
Three funded startups in one month — GridCARE (National Grid partnership), Emerald AI ($68M total), Soma Energy ($7M seed) — all attacking the same problem: finding and dispatching spare grid capacity using AI. The business model is SaaS + utility partnerships, not capital-intensive buildout. For VCs, this is a familiar software playbook applied to a $500B infrastructure bottleneck.

**Project Finance: LDES Graduates from Pilot to Pipeline**
The driver behind long-duration storage's moment: 1) data centers need 24/7 firm power but can't wait for grid connections, 2) regulations in some states now allow industrial loads to proceed with self-supplied backup, 3) lithium-ion can't economically serve 8-100 hour durations. Form Energy's $20/kWh target at 100-hour discharge fundamentally changes the math. EnerVenue's 30,000-cycle nickel-hydrogen adds another bankable chemistry. As offtakes formalize, expect project finance structures to follow.

**Credit: Fuel Cells as Bridge Infrastructure**
Bloom Energy's $8B combined pipeline (AEP + Brookfield) is essentially long-dated power purchase agreements backed by utility-grade counterparties. For infrastructure debt funds, these are contracted cash flows with known technology risk. The investable angle: fuel cells fill the 2-5 year gap before grid connections or nuclear come online.

---

## Seed/A Watchlist

**Astro (YC W26) — Grid Interconnection Intelligence**
Founded by Alex Fuster, former Citadel energy trader. AI models identify profitable grid connections that eliminate the surprise costs causing 80-90% of renewable projects to fail. Addresses the same interconnection bottleneck from the developer side rather than the utility side.

---

## Sources

Fortune, Utility Dive, C&EN (American Chemical Society), Axios Pro, GeekWire, Bloomberg, Energy Storage News, TechCrunch, Crunchbase, Wood Mackenzie, DOE.gov, Y Combinator, Latitude Media