I. Intro

The market looked at NUAI’s recent financing package and focused on the obvious negatives first. More shares. More warrants. More complexity. More dilution.

Fair enough. The cap table clearly got worse.

But I think that reaction misses the more important point: the project got more real. After working through the latest filings, public commentary, and my own modeling, I think the bigger shift is not just that TCDC is now financed in the short term and more financeable in the long term. It is that the economics of the contribution may be materially better than many investors, myself included, were previously underwriting. Several important inputs in the model have changed, and the net effect is meaningfully positive even after accounting for the heavier dilution.

II. Why my view changed

Three things changed my thinking.

First, I no longer think it makes sense to treat the $100 million ATM as the base-case source of near-term funding. The company already did the painful part. It raised the $100 million underwritten equity deal, put the Macquarie structure in place, and based on both the filings and company commentary, the ATM now looks much more like optionality than the default funding path. Management appears to view it as a strategic tool that could be used opportunistically if the stock price is higher or if a compelling growth opportunity arises, not as a blunt instrument to fund day-to-day development or existing TCDC economics.

Second, recent commentary increased my confidence that the site contribution will be valued much closer to powered land than to raw acreage. Raw land at $300,000 per acre is one thing. Powered land with a credible path to commercial deployment, priced with per-MW premiums for power positioning and shell readiness, is something very different. In the model, that distinction flows directly through the JV waterfall into NUAI’s credited contribution and ultimately its common equity ownership percentage. I still think that is directionally right, even if the exact treatment of debt-funded project spend inside the eventual JV remains an important open sensitivity.

Third, the power path for the first 200 MW now looks far more concrete. Before the recent financing, the initial power story still felt relatively vague. The Macquarie loan documents change that picture materially. I walk through the specifics later in this piece, but the short version is that the lender appears to be underwriting toward a customer-specific, site-dedicated power arrangement for the first major block of capacity, not just a generic future grid solution with a massive backlog.

The story is no longer just “interesting site, interesting partner, interesting capital stack.” It is starting to take shape as a real project with real underwritable economics.

III. How I am underwriting TCDC

The model underwrites the full 1,008 MW campus as a single project outcome. I still think that is the right structure as it is the only way to gauge the full potential for TCDC and what that may be worth to NUAI shareholders today.

Total project capex is $13 billion, funded at an 80/20 debt-to-equity split, producing $2.6 billion of total project equity. NUAI’s credited JV contribution, which includes land value plus a portion of debt-funded project advancement, determines its common equity ownership percentage of that total equity pool. Revenue, NOI, and terminal value are all computed at the full campus level, and the JV waterfall distributes proceeds through a preferred return stack and GP promote structure to arrive at NUAI’s share of exit value. The key output is NUAI’s common equity percentage, which in the current base case comes to roughly 30.8%. That figure is higher than before for two reasons. I now assign a full powered-land valuation to the first 200 MW rather than raw acreage, and I model the Macquarie debt being used to fund site improvements that increase NUAI’s equity share in the JV.

IV. Why the contribution value matters so much

This is still the part of the model that matters most.

The JV waterfall breaks the site valuation into three components: a raw land floor at $300,000 per acre across 492 total acres, a Phase 1 powered-land and readiness premium priced per IT MW, and a Phase 2 readiness premium haircut to 35% of the Phase 1 factor. That layered approach captures the fact that NUAI is not handing over dirt. It is contributing a powered, partially development-ready site into the JV, and the value of that contribution should reflect the positioning work that has already been done.

In the current base case, the raw land floor is roughly $148 million. The Phase 1 readiness premium adds about $214 million. Phase 2 adds another $169 million. Total implied site contribution value comes to approximately $531 million. On top of that, the model currently credits $270 million of debt-funded project contribution from the Macquarie facility, bringing total NUAI credited JV contribution to roughly $801 million. That $801 million against $2.6 billion of total project equity is what produces the 30.8% common equity ownership.

My takeaway from the broader financing structure and outside conversations is that the Macquarie facility is best understood as bridge financing meant to start construction, fund long-lead procurement, and add value to TCDC while definitive customer and JV milestones are being worked through. The draw structure is aligned with project milestones, and the capital is best thought of as earmarked for TCDC rather than generic corporate uses. It is a three-year term, which implies that at some point this facility would be transitioned into something more appropriate for a pre-leased hyperscale infrastructure asset once the JV and customer lease are in place. The institutional capital partner would most likely arrange that longer-term project-level debt.

But I did not come away from my conversations with enough confidence to say the market should assume every dollar of that facility gets credited into the JV one-for-one. What those conversations did reinforce is the broader logic: capital deployed into the site should help re-rate the asset and strengthen NUAI’s eventual economics. The exact crediting mechanics still remain a major model sensitivity.

V. Why Stream matters more now

The most important thing about the Stream relationship is how it apparently came together. According to public statements from management, the partner selection was driven by the hyperscaler, not by NUAI. PDI does not appear to have an ongoing role in the Stream JV as currently framed.

If the hyperscaler effectively told NUAI to bring in Stream as the development and operating partner, that changes the signal considerably. It means the customer has already done its own diligence on Stream and concluded that this is the execution stack it wants attached to the site. That is a very different risk profile than NUAI independently picking a partner and asking the market to handicap whether the choice was the right one.

It also has implications for deal probability. A counterparty that is actively shaping the partner stack around a site is not casually exploring options. That is the behavior of a buyer that is trying to get the project to a place where it can say yes, not a buyer that is still deciding whether to engage.

I would not rule out PDI being involved in future projects, but for TCDC specifically, PDI is out and Stream is in.

VI. What the Macquarie documents seem to be telling us

The loan documents appear to contemplate multiple major agreements, not just one headline lease. The “Material Documents” framework includes the Letter of Intent and JV Term Sheet, the definitive JV Agreement, the Data Center Lease, and the Energy Service Agreement. The project is not moving toward one binary customer contract. It is moving through a broader documentation stack that covers governance, customer commercialization, and power delivery.

The cleanest way to think about it is as two tracks that eventually converge.

The first is the JV track. That starts with the March 24 non-binding LOI, which the company later described publicly as a framework with Stream and an institutional investor. Attached to that is the JV Term Sheet, which likely lays out the basic economic and governance structure. The definitive version would be the JV Agreement.

There is one nuance in the Macquarie documents worth flagging. Credit to @kamikazzzi1981 on X for surfacing this logic. The Macquarie term loan defines the “Letter of Intent” as a document dated March 24, 2026, “by and among New Era and the other parties thereto (including the ‘Letter of Intent Party’).” It then separately defines an “Acceptable Hyperscale Entity” as a U.S.-based investment-grade hyperscaler and adds: “for the avoidance of doubt, for purposes hereof, the Letter of Intent Party shall be an Acceptable Hyperscale Entity.” One of the closing conditions also required “an executed copy of the Letter of Intent attaching the JV Term Sheet.” This is not clean enough to prove that NUAI signed a standalone hyperscaler LOI on March 24. But it is enough to conclude that hyperscaler participation was likely already embedded somewhere in the March 24 documentation stack, whether through a direct LOI or a broader multi-party framework. The filings point to something more advanced than a simple Stream-plus-capital-partner narrative, even if the exact structure remains unconfirmed.

The second is the customer commercialization track. That appears to include the Data Center Lease and the Energy Service Agreement. The lease is the definitive customer contract. The ESA is the site-level power arrangement tied to at least 200 MW. More specifically, the Macquarie documents define the ESA as a future agreement between the hyperscaler and the borrower or JV for at least 200 MW of power that may only be delivered to the TCDC properties. That does not mean the ESA is signed. But the lender is clearly underwriting toward a site-dedicated, customer-facing power arrangement for the first major block of capacity, not a generic future utility interconnection with years of backlog risk. The power contract is embedded in the core commercial architecture of the project.

More broadly, the documents read like a contractual stack being assembled around a real development, not a placeholder framework waiting for something to happen.

A few weeks ago, it was easier to frame TCDC as an interesting site that still needed almost everything to come together. Today, the relevant parties appear to be working through a real documentation process with identifiable milestones and defined counterparties. That does not eliminate execution risk. It does mean the project is further along than the market is pricing.

I would not overstate what the filings prove. The Macquarie definitions are not enough to declare a definitive hyperscaler lease is signed, and my channel checks did not add clarity on that point. The right posture is to treat the documentation stack as evidence that the project is real and progressing.

VII. The fee stack adds real value

Based on my own analysis and comparable deal structures, I now have higher confidence that the TCDC economics extend beyond pure land contribution and common equity participation.

The rough framework based on comparable HPC and real estate JV structures appears to include development fees of roughly 2% to 3% of a given phase’s capital during the design and construction period, plus ongoing asset management fees that could look something like 1.25% to 1.5% of LP equity or AUM, depending on how the final structure is set up. Think Digital Bridge as a rough analog for the asset management fee structure. On top of that, there could eventually be some pass-through economics around power from the BTM setup, though I would assign much less confidence to that last bucket at this stage. I am modeling that these fee streams would be split 50/50 between the co-GPs.

I now capitalize a simple version of these fee streams in the base case. Using midpoint assumptions of roughly 2.5% development fees, roughly 1.375% asset management fees, and a 50/50 split with Stream, the fee stack adds approximately $151 million of NPV to NUAI in the model (about 5% of the total NPV). That is real value, but it does not change the core story. Fees are still a modest part of the overall economics. The vast majority of modeled value still comes from NUAI’s common equity ownership and promote participation in TCDC.

VIII. What the dilution actually looks like

The dilution build is worth walking through because it is now meaningfully larger than before.

The starting point is 93.5 million shares issued and outstanding as of April 10, 2026. That figure already reflects the underwritten equity raise as well as additional shares issued to SharonAI as a true-up and to Zachary Zhou on note maturity at the closing of the qualified equity financing. From there, the remaining dilution comes from executive RSUs and PSUs, SPAC warrants, the Macquarie common stock purchase, Macquarie warrants, the 15% underwriter over-allotment, ATW warrants, and the ATM.

There is one final moving piece tied to SharonAI. NUAI has elected to prepay the SharonAI convertible note on April 24, but SharonAI can still elect to convert up to 20% of that note into stock by April 17. The final cash-versus-stock settlement is not fully locked yet.

I no longer include the full ATM in diluted shares. As I discussed earlier, management commentary suggests the ATM is being treated as strategic optionality rather than a primary funding tool, and the broader financing package supports that read. I still flag it as a risk to the share count if circumstances change, but it is not in the base case.

The early trading action also looked better than many people probably expected. The offering settled on April 10, and despite the lack of an obvious immediate lock-up effect for new buyers, the stock did not unravel right after the new paper hit the market. I would not go so far as to say that proves the entire book is made up of long-term institutional holders, but it does suggest the allocation may have been stronger and less flipper-heavy than the market initially feared. That is a constructive signal for a deal this size at this stage of the company’s development.

IX. The next milestones to watch

Outside of the big milestones (finalized JV, official hyperscaler lease, resolution of the NM lawsuit) the near-term catalyst path is more granular than binary. Investors should be watching for a steady stream of smaller LOIs, groundbreaking activity, long-lead equipment orders, and continued progress on the documentation stack. None of those individually move the stock the way a definitive lease would. But they would collectively reinforce the narrative that the project is progressing on schedule and that the company is executing against its timeline.

X. Risk

The main thing that can still go wrong is simple: the hyperscaler does not sign. That is not impossible, and it has happened in this sector before where hyperscalers walk away at the altar. Even if TCDC would still likely attract interest over time, the consequence would be a slower timeline, weaker confidence, and probably less attractive economics.

That remains the key risk. Everything else is secondary.

XI. Bottom line

A few weeks ago, the bear case was that TCDC might never become a real project. That framing no longer holds. The company raised $100 million in equity at a meaningful but not obviously distressed discount to market, put a $290 million credit facility in place with milestone-aligned draws, and the Macquarie documentation stack reads like it was structured around a project with an identifiable hyperscaler counterparty.

My model now produces a probability-weighted price target of roughly $19.27 per share, or roughly $24.09 at 100% success.

What matters most from here is simple. The hyperscaler either signs or it does not. If it signs, the powered-land contribution framing likely holds, the JV gets documented, and the equity re-rates. If it does not, the timeline stretches, the economics get worse, and the capital structure becomes a heavier burden. There is no middle path where the stock works without a real customer.

The cap table got worse; the project got more real. I know which one I am betting on.

View my latest NUAI model here.

Updated Conservative Base Case PT: $8.62

Updated Realistic Base Case PT: $19.27

Updated Bull Case PT: $37.17

Subscribe now

Share

Disclosure: I am long NUAI. I may buy, sell, or change my position at any time without notice. This article reflects my personal opinions, is for informational and educational purposes only, and is not investment advice or a recommendation to buy or sell any security. I have not received compensation from NUAI, its management, or any affiliated party for this content. I have no business relationship with the company.

• This model values TCDC only as a shell campus JV, not a full stack AI infrastructure play and not a broader NUAI platform case.

• Under the Realistic Base Case, the model produces a $21.33 per share 100% success case and a $16.00 probability weighted value, driven by $1.679 billion of 2030 NOI, a 6.0% terminal cap rate, a 1.35 PUE assumption, and 10.06% common ownership for NUAI.

• The Stream Data Centers LOI moves TCDC closer to the standard private capital playbook for hyperscale development: originator, specialist operator, and institutional capital. Apollo is the most natural inferred candidate for the unnamed capital partner.

• The land contribution build is the heart of the model. It is not estimating what the land would fetch in a spot sale. It is estimating what a JV counterparty might reasonably credit NUAI for contributing the site. That framework is basis adjusted to the same IT load logic used in the operating build. After explicit site capture and readiness haircuts, that results in $243.54 million of credited value, 10.06% common ownership in the JV, and a 50% share of the GP promote pool.

• The model deliberately excludes several things that could inflate the target, including the potential 7 GW New Mexico development, Phase 3 readiness value, behind the meter generation economics, undefined fee streams, and any speculative platform overlay beyond TCDC.

NUAI Model v16:

View Updated Full Model Here

Intro

Most investors looking at TCDC seem to fall into one of two camps. One sees a large AI data center opportunity and immediately jumps to a giant upside number, often by applying a very rough APLD comp to total MW capacity, which I think is the wrong framework for several reasons. The other sees a project this large, this early, and this dependent on future agreements and concludes it can’t be modeled seriously at all.

I think both approaches miss the real question.

The relevant question is narrower: if TCDC becomes the shell campus for a Big 4 hyperscaler that management has been describing, what’s that worth to NUAI on a fully diluted basis?

That’s what this model is trying to answer. It isn’t a full stack AI infrastructure model, and it isn’t assigning value to every dollar of behind the meter generation, tenant equipment, or distant optionality. It’s a shell campus JV model. The framework is simple: estimate what the campus could be worth operationally, estimate what value a JV would likely credit to NUAI for contributing the site, run the JV waterfall, and then translate NUAI’s share into per share value on a fully diluted basis.

The model converts gross campus MW into revenue generating IT MW through an explicit PUE assumption before applying lease comps, and it basis adjusts the land contribution framework the same way.

Under the current Realistic Base Case, that produces a 100% success case value of about $21.33 per share and a probability weighted value of about $16.00. The real question, however, is whether the assumptions behind those outputs are defensible.

What The Model Is Actually Valuing

The most important choice in the workbook is the valuation perimeter. This is a shell campus JV model, not a model of the full AI infrastructure stack that could eventually be built on the site. The valuation is based on the economics of the shell campus SPV that NUAI is trying to participate in, rather than the full economic value of every power asset, tenant fit out, compute layer, or other infrastructure component that may ultimately sit at TCDC.

That distinction matters because it changes both the capital required and the portion of value that should reasonably be credited to NUAI. If NUAI were modeled as though it were funding and owning the entire stack, the analysis would overstate both CapEx and NUAI’s economic claim on the project. If it were modeled as though it only controlled raw acreage, the analysis would understate what a real JV counterparty might reasonably credit for a site with meaningful development and strategic value.

The model therefore breaks the problem into three parts. First, it builds the operating case for the shell campus itself, using an explicit conversion from gross campus MW to revenue generating IT MW. Second, it estimates what a JV counterparty might credit NUAI for contributing the site. Third, it determines how much of the resulting value NUAI would receive through common ownership and sponsor economics. Put differently, this is neither a raw land appraisal nor a full stack infrastructure model. It is a valuation of NUAI’s likely economic participation in the shell campus JV.

A few terms are worth defining at the outset. Here, a NNN shell lease means the landlord is primarily delivering the site, shell, and core, while the tenant bears or reimburses most of the operating burden. LTC, or loan to cost, is debt as a percentage of total project cost. The cap rate is the rate used to value the property at exit. The promote is the sponsor’s carried interest style share of residual value after capital and preferred return have been paid first.

What Is More Market Backed And What Is More Judgment Based

Before walking through the mechanics, it helps to be clear upfront about which inputs are grounded mainly in market data and comparable analysis and which rely more heavily on informed judgment. That gives the reader a better sense of how much weight to place on each assumption as it appears.

The more market backed inputs are the anchor and premium rent range, the residual owner OpEx framework, the shell CapEx scope, the 80% LTC assumption and broader interest rate environment, and the terminal cap rate range for a successful hyperscale end state.

The more judgment based inputs are the hyperscaler anchor and premium tenant mix, the PUE and resulting conversion from gross campus MW to revenue generating IT MW, probability of success, site capture, Phase 2 readiness, shell and readiness uplift, the final promote split, and the ownership credit NUAI would ultimately receive at close.

The model would be weaker if it tried to present every major assumption as though it were equally observable or equally certain. Some inputs are anchored primarily in market data. Others are closer to judgment. Both matter, but they shouldn’t be treated equally.

The Realistic Base Case

The Realistic Base Case assumes anchor hyperscaler shell rent of $170 per kW per month, premium AI tenant rent of $255 per kW per month, likely from a neocloud tenant, a 70% anchor and 30% premium end state mix, a 1.35 PUE, residual owner OpEx of $75,000 per IT MW, a 6.0% terminal cap rate, 7.0% interest, 80% LTC, a 9.0% discount rate, and a 75% probability of success.

On those assumptions, the model produces a site contribution value of $243.54 million, NUAI common ownership of 10.06%, 2030 NOI of $1.679 billion, terminal enterprise value of $27.98 billion, equity value at sale of $20.38 billion, NPV of NUAI common equity of $1.132 billion, NPV of NUAI promote of $1.138 billion, total NPV to NUAI of $2.270 billion, a fully diluted share count of 106.42 million, a 100% success case value of $21.33 per share, and a probability weighted value of $16.00 per share.

Those outputs are only as credible as the assumptions driving them.

What The Model Is Assuming Mechanically

Before getting into the operating assumptions, it’s worth being explicit about how the workbook actually translates those assumptions into value.

First, this is fundamentally an exit value model. The equity value is realized primarily through a 2030 terminal valuation rather than through interim cash distributions. During the build and ramp period, the model doesn’t assume meaningful current yield to equity. It assumes the project is built, stabilized, and then monetized off 2030 NOI.

Second, annual NOI is effectively swept to debt paydown before exit. In the model, debt repayment each year is set equal to annual NOI. That means there are no interim distributions to equity, no separate reserve build, and no explicit split between required amortization and optional deleveraging. The result is that ending debt at the 2030 exit is approximately $7.60 billion, which means equity at sale is the difference between the $27.98 billion terminal enterprise value and that remaining debt balance.

Third, the terminal value is based directly on 2030 NOI and a cap rate. There’s no explicit post 2030 operating forecast, no multi year hold beyond that point, and no lease mark to market analysis beyond the 2030 stabilized year. The cap rate is doing the work of converting one year of modeled NOI into long duration asset value.

Fourth, the model no longer applies lease comps directly to the full 1,008 MW campus. It first converts gross campus MW into revenue generating IT MW through an explicit PUE assumption. In the Realistic Base Case, that PUE is 1.35, which means the 1,008 MW campus translates into about 739.2 revenue generating IT MW at full buildout.

Fifth, the model assumes energized capacity is monetized almost immediately. Utilization is 99% in every operating year once a phase is online. So the commercial ramp is driven mainly by when MW are energized and by rent mix, not by a drawn out lease up curve.

Sixth, this is a property level shell NOI model, not a corporate free cash flow model. It doesn’t attempt to capture corporate SG&A, public company overhead, cash taxes, reserve accounts, transaction costs, leasing commissions, or other platform level leakage that would sit outside a shell SPV property valuation framework.

Seventh, the fully diluted share count includes both visible dilution and an analytical dilution buffer. The per share output uses 106.4 million shares, including basic shares, SharonAI related dilution, executive PSUs, and an additional shelf buffer that serves as a modeled conservatism layer rather than a fixed disclosed endpoint.

Finally, NUAI’s site contribution isn’t treated as a single day one drop into the JV. In the model, it’s allocated against the annual project equity draw schedule. That means contribution timing affects ownership economics and waterfall claims, rather than simply converting land value into a flat upfront percentage with no regard for when the capital is actually needed.

The Revenue Build

The revenue build uses two rent buckets: a Big 4 hyperscaler anchor shell rent and a higher rate premium AI tenant rent (likely a neocloud). In the current Realistic Base Case, the anchor tranche is $170 per kW per month, the premium tranche is $255, and the end state mix is 70% anchor and 30% premium. That implies a stabilized blended rent of $195.50 per kW per month, or about $2.356 million per IT MW per year.

The annual ramp is also modeled more realistically than in a simple blended rent build. The anchor allocation fills first, and the premium tranche only begins to appear once the anchor bucket is filled. That’s a better approximation of how lease up would likely work in practice. If the end state really is a Big 4 style anchor led shell campus, the early signed MW should resemble anchor hyperscaler economics, not a perfectly blended rent mix from day one.

One thing that isn’t happening in the revenue build is any reliance on contractual annual rent escalators. Revenue growth is coming from the MW ramp and the shift in mix between anchor and premium tranches, not from embedded annual price step ups. That keeps the build cleaner and adds some conservatism, since the 2030 NOI is being driven primarily by stabilized capacity and mix rather than by contractual escalation over time.

No single comp will give you the exact right answer, but the range is grounded in market evidence. Hut 8 / Fluidstack / River Bend implies roughly $159 per kW per month. Applied Digital / CoreWeave screens around $156, while Applied Digital’s investment grade hyperscaler lease implies something closer to $139. At the broader market level, CBRE’s H2 2025 U.S. wholesale colocation asking rates were about $196.25 per kW per month. Iron Mountain’s disclosed leasing prints, roughly $202 on new and expansion deals and $297 on renewals, are more useful as a check on the premium tranche than as a benchmark for the entire campus. On that comp set, I don’t think a $170 anchor rate is aggressive, particularly since the $255 premium rate is only applied to 30% of end state economics rather than the full project.

From Rent Per MW To 2030 NOI

The math from rent to NOI is still straightforward, but the MW basis has changed. At a stabilized revenue density of roughly $2.356 million per IT MW, and with about 739.2 revenue generating IT MW at full buildout under the 1.35 PUE assumption, the model produces about $1.734 billion of 2030 revenue. The model then subtracts residual owner OpEx of $75,000 per IT MW, or about $55.4 million in aggregate, leaving 2030 NOI of about $1.679 billion.

That OpEx number is low relative to the full cost of operating a completed data center, but that is not what the model is capturing. This is residual landlord level OpEx in a shell heavy, NNN style structure where the tenant bears or reimburses most of the actual operating burden inside the asset. In other words, the model is not assuming a data center can be fully operated for $75,000 per IT MW. It is assuming the landlord retains something in that range after lease structure and pass throughs shift most of the cost burden to the tenant.

That distinction matters because this is the core bridge from operating assumptions to value. The large terminal valuation later in the workbook is coming from capitalizing a very large NOI stream, not from hidden assumptions elsewhere in the model. And because this is an exit value model, that 2030 NOI line is doing more work than it would in a longer hold DCF, which makes the residual OpEx assumption one of the central underwriting variables in the workbook.

CapEx And Leverage

The model assumes roughly $12.1 billion of total CapEx inside the shell SPV, made up of about $11.5 billion of shell build, $450 million of soft costs, and $150 million of land and site prep. Across 1,008 MW, that comes out to roughly $12.0 million per MW.

JLL’s shell and core cost discussion around $11.3 million per MW is a useful directional check and is broadly consistent with the CapEx level used here.

The leverage assumption is one of the larger drivers in the workbook. The model uses 80% LTC, which has a major effect on equity value, but it is not a made up assumption. Management has explicitly pointed to roughly 80% LTC multiple times, including in connection with the Stream LOI, and the broader comp set is directionally supportive. Galaxy’s Helios Phase I financing screens around 82% LTC, Hut 8 has discussed leverage up to 85% LTC, and recent Starwood construction loan disclosures support the general rate environment being used. In other words, the model is not stretching to an 80% LTC structure. It is using the project level capital structure management has been directly signaling.

One distinction worth noting is that the CapEx build remains on a gross campus MW basis, while the revenue build in the model is now converted to revenue generating IT MW through the PUE assumption. That changes the operating side of the model, but it does not change the shell campus CapEx framework itself.

It’s also important to distinguish between starting leverage and ending debt at exit. Because the model monetizes value through a 2030 sale, equity value depends not just on the initial 80% LTC, but on how much debt remains outstanding at that point. In the workbook, that ending balance is now approximately $7.60 billion and is driven by annual NOI sweeps and accrued PIK interest over the life of the project, not simply by the original leverage assumption.

Why The Stream Partnership Matters

The April 1, 2026 LOI with Stream Data Centers is the most important recent development in the story. If it progresses to definitive documentation, it changes the underwriting framework materially.

The LOI describes a far more institutional structure than what many investors had likely been assuming. NUAI contributes site control and local relationships. Stream is expected to serve as development manager and operator. An unnamed Institutional Investor contributes equity capital and leads project financing, which is expected to include roughly 80% debt financing on market terms. That’s a familiar structure in digital infrastructure: the local originator brings the asset, the specialist operator brings development execution and customer credibility, and the capital partner brings the balance sheet.

That matters because TCDC doesn’t just need capital. It needs counterparties that both hyperscalers and lenders will take seriously. Stream helps address that problem. Stream isn’t a first time developer trying to enter the category. It describes itself as a hyperscale and enterprise data center developer and operator founded in 1999, with more than 27 developments delivered, over 4 GW of pipeline, and over 90% of capacity leased to Fortune 100 customers. It has also said that all of its projects have been delivered as joint ventures, which is especially relevant here because it suggests the platform already knows how to operate in multi party capital structures rather than needing to learn that on the fly.

More specifically, Stream matters here because TCDC does not simply need a contractor. It needs a platform that can help turn a large, early stage site into a deliverable hyperscale project that is credible to customers, lenders, and institutional capital. That requires development execution, leasing credibility, operating capability, and experience working inside JV structures. Stream brings those capabilities in a way a local sponsor alone typically cannot.

The LOI doesn’t identify the Institutional Investor, so I don’t want to overstate what’s disclosed. But the Apollo connection still matters even if Apollo isn’t named directly in this project. Apollo acquired a majority interest in Stream in 2025 and framed that deal around providing long duration capital to help accelerate Stream’s multi gigawatt pipeline. That doesn’t prove Apollo is the exact capital partner here. But it does mean Stream is operating from a very different position than a standalone private developer that still needs to raise capital project by project. At minimum, Apollo is the clearest reason the market should view Stream as a much more serious financing counterparty than an ordinary private developer.

If investors are asking who the most likely candidate for the unnamed capital provider is based on the public facts, Apollo is the most natural place to look. The logic is straightforward. Apollo bought control of Stream specifically to back a multi gigawatt digital infrastructure platform, and the TCDC structure outlined in the LOI fits that model closely: an originated site, a specialist operator, and a large institutional capital partner funding the asset with a mix of equity and heavy asset level leverage. That remains an inference rather than a disclosed fact, but it is a well grounded one.

Why A 50/50 GP Split Is A Reasonable Working Assumption

The exact sponsor split is not public, and the 50/50 co GP concept in the model remains an analytical assumption rather than a disclosed term. I still think it’s a reasonable starting point, but the right way to frame it isn’t simply that there are three parties involved.

The more important question is what role NUAI is actually playing in the structure. If NUAI were effectively just selling land into a larger JV and stepping back, then a 50/50 sponsor split with Stream would likely be too generous. In that case, the site contribution would matter, but most of the ongoing sponsor economics would more naturally accrue to the operating partner and capital provider. On the other hand, if NUAI is being treated as a real originating sponsor that sourced the site, assembled the local position, advanced development work, and remains central to the project rather than just monetizing land, then a shared co GP framework becomes much more logical.

That’s why I think 50/50 is a reasonable base assumption today. The public structure suggests NUAI is not simply exiting into the deal, and Stream is not merely acting as a third party consultant. NUAI appears to be contributing the site and local development foundation, while Stream is contributing execution, leasing, and operating capability. In that setup, a shared sponsor economics arrangement is directionally consistent with how large private capital backed developments are often structured, even if the final documents may split particular fees, promote buckets, or governance rights differently.

How PE Backed Data Center Builders Typically Work

More broadly, this is how large private capital backed data center development increasingly works. The local or originating party brings the site, development groundwork, and early project creation. The specialist operator brings execution, customer relationships, and credibility with lenders and hyperscalers. The institutional capital partner brings the equity check and supports a highly levered project financing structure at the asset level. The reason these deals are structured this way is simple: hyperscale campuses are too large and too capital intensive to fund cleanly off one balance sheet unless the platform is already enormous.

That’s the broader significance of the Stream relationship. It potentially moves TCDC out of the category of “small company with a large site” and into the category of “originated site being pulled into a scaled private capital development machine.” In practice, that usually means more institutional underwriting, more credible development execution, better access to financing, and a much higher probability that the site is evaluated as a real deliverable campus rather than just a speculative story.

For TCDC specifically, I think that changes the way investors should think about risk. It doesn’t eliminate risk, and it doesn’t tell us the final economics. But it does suggest the project is moving closer to the standard private capital playbook for large data center development: originator plus operator plus institutional capital, with sponsor economics shared across the parties actually creating value.

Terminal Cap Rate, Discount Rate, And Probability

The Realistic Base Case uses a 6.0% terminal cap rate, a 9.0% discount rate, and a 75% probability of success.

With $1.679 billion of 2030 NOI and a 6.0% cap rate, the model arrives at roughly $27.98 billion of terminal enterprise value. I think 6.0% is a reasonable base case exit multiple if the project reaches a successful hyperscale end state. The model isn’t assuming a generic speculative shell development at exit. It’s assuming a leased, institutional quality campus. If that happens, the exit should be underwritten more like stabilized, long duration digital infrastructure than like speculative development land.

The 9.0% discount rate is doing something different. It’s still meaningfully above what one might use for fully de risked stabilized infrastructure because the model is discounting a success case stream that still has execution, timing, financing, and documentation risk in front of it. The 75% probability haircut is meant to answer a separate question: what are the odds that the project actually reaches that signed, financed, stabilized end state at all.

There is some conceptual overlap between discounting for risk and applying an explicit probability weighting, and reasonable investors can disagree on the right treatment. I prefer to show both separately rather than force all of that work into a single rate. In this model, the cap rate values the exit, the discount rate reflects time and risk within the success case, and the probability haircut reflects the likelihood that the success case is realized at all.

The Land Contribution Build

This is the most important section in the model because it bridges what NUAI controls today to the common equity ownership it is ultimately credited with in the JV. More than any other part of the workbook, it determines whether the implied valuation rests on a credible underwriting framework or on an overly generous view of the site contribution.

The model isn’t trying to produce a literal spot market appraisal for the site. It’s trying to estimate what a JV counterparty might reasonably credit to NUAI for contributing the site. That’s a different exercise. A raw land only approach would likely understate the negotiation value of the asset. A valuation based on fully matured shell campus economics would clearly overstate it. The model therefore builds a middle ground with three components: a raw land floor, a powered land premium, and an additional shell and readiness uplift.

The raw land floor is straightforward. The model credits 438 owned acres at about $244,000 per acre, producing approximately $106.872 million. That provides a real floor without requiring aggressive assumptions.

The next layer shifts from a dollars per acre framework to a dollars per MW framework. That is intentional. Once the scarce input is no longer land itself but power and deployable capacity, acreage alone becomes a less useful valuation lens. The model applies a powered land premium and a separate shell and readiness uplift on the same IT-load basis used in the operating build, then basis adjusts those values through the PUE assumption before applying them to the campus phases. That is important because it keeps the site contribution framework internally consistent with the revenue build rather than mixing IT-load-based comps with gross campus MW.

I do not view recent powered site transactions as fixing the exact right number, but they do provide a reasonable valuation frame. Bit Digital’s NC-1 amendment reset the purchase price to $45 million for a site with 99 MW contractually targeted within four years, with management also discussing potential for up to 200 MW over time. Digital Realty’s Atlanta acquisition was $120 million for a site expected to support more than 200 MW, or roughly $600,000 per MW. The model uses those kinds of signals directionally rather than treating any one comp as definitive.

The third layer remains the most judgment heavy. The shell and readiness uplift is not meant to imply the land itself could be sold today for that incremental value in a plain vanilla transaction. It is meant to reflect the possibility that a site is worth materially more to a JV counterparty than raw land plus basic power relevance if it improves speed to market, reduces development friction, and helps make the project more financeable.

The important point is that the model does not give NUAI full credit for the gross benchmark. It applies two explicit haircuts: a site capture haircut and a readiness haircut. In the current Realistic Base Case, site capture is 40%, Phase 1 is 200 MW at 75% readiness, and Phase 2 is 450 MW at 35% readiness. After the basis adjustment and those explicit haircuts, that results in Phase 1 credited premium of about $66.67 million and Phase 2 credited premium of about $70.00 million. Put differently, the model is not giving NUAI full credit for future shell economics today. The haircut structure materially reduces the amount of premium value recognized at close.

One more detail that matters here: the workbook includes 54 pending corridor acres, but corridor inclusion is currently set at 0% in the Realistic Base Case. So the base case land contribution value isn’t quietly relying on speculative corridor acreage. It’s coming from owned acreage plus the powered and readiness framework attached to Phase 1 and Phase 2 MW.

Put together, the three pieces produce a final contribution value of $243.54 million: $106.872 million of raw land floor, $66.67 million of Phase 1 premium, and $70.00 million of Phase 2 premium. Because the model assumes 80% LTC, total project equity is about $2.42 billion. That converts into roughly 10.06% common ownership.

I don’t think the site would necessarily appraise today for exactly $243.54 million in a clean sale process. I do think it is a reasonable framework for estimating what a JV counterparty might credit NUAI for contributing the site. The model gives full credit to raw land, basis adjusts the powered land and shell/readiness framework to the same IT-load logic used elsewhere in the model, isolates the most judgment heavy readiness component rather than obscuring it, and then applies explicit haircuts so the credited value does not amount to giving NUAI full future shell economics at close.

From Land Value To Ownership

Once the contribution value is built, the model translates it into common ownership. That is why the front page sensitivity tables use NUAI common ownership rather than raw land value itself. Ownership is the cleaner reduced form output of the entire land contribution debate. Once the site contribution is converted into 10.06% common ownership, a reader can see the valuation impact directly without rebuilding every acre and MW assumption underneath it.

One additional nuance matters. In the workbook, NUAI’s site contribution is allocated against annual equity draws rather than being treated as a single day one contribution. That means ownership economics and waterfall claims are influenced not just by contribution value, but also by contribution timing.

The JV Waterfall

Once the model has project equity value at exit and NUAI’s common ownership, it runs the JV waterfall. The sequence is straightforward: return contributed capital, pay the preferred return, determine the residual above the hurdle, carve out the promote pool, and allocate the remainder to common equity.

The model uses an 8% hurdle, a 20% promote above that hurdle, and a 50% NUAI share of the GP promote. That means NUAI’s effective promote share is 10% of the residual above the hurdle. In the current Realistic Base Case, capital plus preferred return amount to about $2.867 billion. Equity value at exit is about $20.379 billion. Residual above the hurdle is therefore about $17.513 billion. The GP promote pool is about $3.503 billion, of which NUAI’s share is about $1.751 billion. At the same time, NUAI receives common equity value based on its 10.06% common ownership, which produces a common distribution at exit of about $1.741 billion.

Discounting those distributions back to today produces an NPV of $1.132 billion for NUAI’s common equity and $1.138 billion for NUAI’s promote. In other words, the promote is not incidental to the valuation. It is a core part of the economics, representing roughly half of the modeled value.

That is why the final JV structure matters. If the promote split, hurdle rate, or sponsor sharing terms end up materially less favorable than the current build assumes, the implied equity value would come down accordingly.

From NUAI Value To Price Target

Once the model has total NPV to NUAI, the price target math is straightforward. In the Realistic Base Case, total NPV to NUAI is about $2.270 billion. Dividing that by a fully diluted share count of 106.419 million produces a 100% success case value of about $21.33 per share. Applying the 75% probability of success results in a probability weighted value of about $16.00 per share.

Using a fully diluted share count matters. Even if TCDC works, NUAI is not reaching that outcome on today’s undiluted base. Some additional parent level dilution is a realistic assumption, mainly to fund G&A and corporate operating needs before project cash flows begin to matter. That is separate from funding the campus itself, which is financed at the SPV level.

The 106.419 million share count includes basic shares, SharonAI related dilution, executive PSUs, and an additional miscellaneous shelf buffer, which is included as a conservatism layer rather than as a disclosed endpoint.

Scenario Analysis And Sensitivities

The model currently produces a Conservative Base of $5.57 probability weighted and $9.28 in a 100% success case, a Realistic Base Case of $16.00 and $21.33, and a Bull Case of $37.96 and $44.65. That range is appropriately wide given the number of variables that remain subject to underwriting judgment. The bull case is not just the base case with slightly improved assumptions. It reflects stronger rent, lower residual OpEx, a tighter terminal cap rate, a lower PUE, a higher probability of success, higher site capture, and greater Phase 2 readiness. The conservative case pushes those variables in the other direction. It is not a failure scenario, but rather a version of the project that works on less favorable economics than the base case.

The front page sensitivity tables focus on terminal cap rate versus NUAI common ownership and probability of success versus NUAI common ownership. That’s the right framing because it allows the reader to pressure test the key value drivers without rebuilding the model: ownership, exit multiple, and success probability. Common ownership sits on the vertical axis deliberately, because it is the most useful reduced form expression of the entire land contribution and JV economics debate.

Mechanically, the model breaks fastest if one of four things is wrong: the gross-campus-MW-to-IT-MW conversion is too aggressive, the site contribution is credited materially below the current build, the project doesn’t achieve near full monetization shortly after energization, or the final capital structure and JV economics come in below the current 80% LTC and 50% promote share assumptions.

The IRR Question

The very high levered equity IRR, currently about 129.25%, should not be read as the core project return. It is a sponsor equity outcome driven by high leverage and a backend loaded contribution profile. The better economic anchor is the unlevered project return, which the model shows at about 43.15% IRR and 2.31x MOIC.

The Energy Dome MOU

The other recent strategic development worth addressing is the April 2, 2026 MOU between NUAI and Energy Dome.

I don’t think this should yet be treated as an explicit modeled input. At this stage, it remains an MOU and an assessment framework rather than a contracted deployment with defined scope, CapEx, and operating economics. Energy Dome would be more likely to affect project economics through efficiency, uptime, and product quality than through a standalone cash flow line, and I do not think there is enough definition yet to underwrite that impact in the base case.

Even so, the strategic logic is worth laying out if the system is ultimately deployed. Energy Dome’s core technology is the CO2 Battery, a long duration energy storage system based on a closed thermo mechanical cycle. At a high level, the system uses electricity to compress CO2 gas into a liquid for storage, then later expands it back into gas through a turbine to generate power. The standard system targets the 8 to 24 hour duration range, uses common industrial materials, and avoids many of the degradation and supply chain issues associated with lithium ion systems.

For TCDC, the more relevant configuration is CO2 Battery Plus rather than the standard storage configuration. In a standard CO2 Battery, the heat generated during compression has to be captured and stored so it can later be used during discharge. CO2 Battery Plus is designed to sit alongside open cycle gas turbines and use their exhaust heat instead. A simple way to think about it is that the system is trying to turn heat that would otherwise be wasted into additional dispatchable power.

That matters because TCDC is not trying to solve a generic renewable grid storage problem. It is trying to solve a more specific data center power problem: getting power online faster, managing the mismatch between jagged AI load profiles and on site generation, and lowering delivered energy cost and carbon intensity without sacrificing uptime. If the system can convert wasted exhaust heat and stored pressure into dispatchable power, it can improve both the flexibility and the efficiency of a gas turbine power island.

There is also some real outside validation here. Energy Dome announced a strategic commercial agreement with Google in 2025, backed by a direct investment. That does not guarantee deployment at TCDC, and it should not be treated as equivalent to a signed project level contract. It does suggest the technology is credible enough to clear diligence with a sophisticated hyperscaler buyer.

Operationally, the value proposition is straightforward. In one mode, the system can use turbine exhaust heat to increase continuous net output without incremental fuel burn. In another, it can help buffer load spikes by converting stored CO2 into temporary surge power. Whether that ultimately improves TCDC’s economics will depend on final design, deployment scope, and system cost. For now, I view it as a strategic positive rather than something that should be explicitly modeled.

What Is Excluded From The Model

One reason I am more comfortable with the model is that it leaves out several things that could easily have been used to push the price target higher, which adds a degree of conservatism to the framework. The model does not include a speculative platform overlay beyond TCDC, any value for future campuses outside TCDC including the potential 7 GW New Mexico development, explicit Phase 3 readiness value in the site contribution, or additional fee streams such as development and management fees that management has discussed in the past. Those items may ultimately prove valuable. I just do not want them doing hidden work inside the headline target before they are sufficiently defined.

What Moves the Needle Most From Here

If I had to rank the remaining variables that matter most, I would put them in roughly this order: first, what Stream Data Centers ultimately credits NUAI for the site contribution; second, how much of the project economics NUAI captures through common ownership relative to sponsor economics; third, whether the final JV waterfall is as favorable as the current promote assumptions; and fourth, how much execution risk should be embedded in the probability of success.

The first question is still the largest driver. If the site is ultimately credited much closer to raw land value, the price target moves down. If it is credited closer to a true development readiness value, the current base case could prove conservative.

Bottom Line

I don’t think the right takeaway from this exercise is that NUAI is worth exactly $16.00 today in some precise mechanical sense. I do think the model makes a narrower point: if TCDC becomes the kind of Big 4 hyperscaler anchored AI campus management has been describing, and if NUAI is ultimately credited something close to the contribution value modeled here, the equity upside can still be substantial even after applying a meaningful haircut for execution risk.

That, to me, is the right way to use the model: not as a claim of precision, but as a transparent framework for underwriting what the upside could look like if the project comes together on something close to these terms.

For a while, the cleanest bear case on AI was pretty simple. Pre-training was supposed to flatten. Models would still get better, just at a slower rate. Efficiency would improve. Cost per token would fall. And eventually the market would realize it had gotten way too excited about an infrastructure buildout chasing demand that would simply fail to materialize.

I think 2026 has made that story much harder to believe.

AI may require far more compute than the market still assumes. The next major gains may depend on giant coherent compute systems that are far harder to build than most people appreciate. And the companies that can actually deliver that infrastructure are positioned to capture an outsized share of value.

The why boils down to three things.

1. The Anthropic Mythos leaks suggest pre-training scaling laws may be better than forecasted. If the leak is even directionally right, Mythos came in dramatically higher in performance than standard scaling laws would have implied.

2. xAI’s Colossus buildout suggests the next major leap may depend not just on more compute in the abstract, but on massive, gigawatt-scale coherent compute. Giant, tightly integrated clusters where power, cooling, networking, and GPUs all have to work as one system.

3. Agents and test-time compute are piling massive incremental inference demand on top of everything else.

The Mythos Leak

For the last year or so, the market got increasingly comfortable with the idea that frontier pre-training was flattening. Bigger runs would still help, but the era of real step-function gains from simply pushing scale harder was supposed to be ending. The implication was simple enough: the next leap would have to come from some architectural breakthrough, clever training tricks, or some genius insight from a frontier lab researcher. Not from brute-force scale itself.

But if these leaks are even directionally right, Mythos points precisely the other way. If a frontier lab can still push a massive run and get a real step change instead of a modest incremental gain, the old diminishing-returns narrative looks much weaker. Some people are starting to theorize that once you cross a certain compute threshold, possibly around the gigawatt scale, progress stops looking linear and starts looking more like a phase transition. The system looks only marginally better for a while, and then once it crosses some critical level, it suddenly becomes considerably more useful.

Think about an airplane. Below a certain speed, adding more power just makes it roll faster on the runway. But once it crosses the lift threshold, it does something fundamentally different. It flies.

If true, this changes the capital allocation logic for hyperscalers and frontier labs. The incentive structure shifts toward continuing to increase compute budgets because the payoff curve is still there. That means bigger training clusters, more capex, and a lot more urgency around time to compute.

Why Colossus Matters

If Mythos is the signal that training scaling laws may still be stronger than people thought, Colossus is the signal that the next step may require a different kind of infrastructure entirely. This is not just about the number of GPUs. It is about how well those GPUs communicate with each other.

I believe that Jensen Huang explains this better than anyone else with his famed AI factory analogy. A warehouse stores things. A factory produces things. If AI were mostly a software layer sitting on top of generic infrastructure, the story would stay in software. Better apps, better interfaces, better distribution. But useful intelligence, measured in tokens, is increasingly something you manufacture through giant integrated systems.

NVIDIA is no longer focused simply on the GPU itself but on the entire compute stack surrounding the chip through what Jensen calls “extreme co-design.” Throughput, utilization, power, cooling, networking, coherency, time to capacity. All of it matters. You are not just plugging in chips. You are building a factory from the ground up, all the way from the silicon to the power infrastructure and the physical shell housing the servers.

Other frontier labs can spend enormous amounts on distributed compute across cloud platforms but that is not the same thing as having one giant coherent training machine. Colossus looks like the first true gigawatt-scale coherent AI training system in the world. It is already operating around 1.0 GW with a path toward 1.5 GW, built around roughly 550,000 Blackwell GPUs, and therefore arguably almost three times larger than any other active training environment.

Once you get to this scale, the bottleneck is no longer simply “buy more chips.” It is communication. Synchronization. Locality. Keeping a massive number of accelerators working together without the whole system bogging down waiting on the network. A recent empirical study on production-scale GPU training clusters found that once workloads move beyond a small number of nodes, network topology, congestion, collective synchronization behavior, and GPU locality can dominate end-to-end training performance. Another survey on distributed training makes the same point more broadly: as computation gets faster, communication becomes a larger share of total training time, making network design increasingly critical.

That is why, for training, coherent compute matters so much more than just renting a ton of distributed cloud capacity. Raw GPU count can tell you how much silicon you own. It does not tell you whether the cluster can actually train at the frontier efficiently. Large distributed training often hits diminishing returns because communication starts eating the gains from adding more workers.

Empirical data from NVIDIA’s Hybrid-EP writeup illustrates the point well. In DeepSeek-V3-style MoE training, communication time can account for more than 50% of total training time without optimization, and NVIDIA’s optimized communication stack delivered 514% throughput improvements in real-world large-scale workloads. This is exactly what the extreme co-design approach is about. You aren’t just buying chips. You’re building a coherent factory.

That is why I think Grok 5 matters so much. It looks like it will be the first model trained on roughly 1 GW of coherent compute. Maybe it disappoints. Obviously that is possible. But if it does not, and if it really is a major step change, then the market is going to have to grapple with a much bigger idea: the next frontier may belong not just to whoever has the best model code, but to whoever can assemble and operate the biggest coherent compute factories. The race to AGI may end up looking a lot more like a systems and manufacturing problem than people still expect.

In other words, atoms may win over bits.

The New Scaling Law

The story does not stop at pre-training. One of the biggest changes in AI over the last year is that scaling is no longer just about training. There is now another axis that matters a lot: test-time compute.

A model does not just get better because it was trained on more data or with more GPUs. It can also get better by spending more compute during inference. More thinking, more verification, more branching, more attempts, more time before it commits to an answer. The old mental model of one prompt in and one answer out is already outdated.

Then agents come into the picture: OpenClaw, Claude Code, Codex, etc.

A chatbot answer can be short and cheap. An agent is different. A task comes in. The model reasons. It searches. It uses tools. It reads files. It checks itself. It may split the work across multiple sub-agents. Then it pulls everything back together and delivers the final answer. The user sees one output. The system may have done far more work under the surface. That is why Jensen’s line that “thinking is way harder than reading” lands so well. It is the simplest explanation of why inference demand can grow much faster than people expect. The cost is no longer just the answer. It is the entire reasoning loop that led to the answer.

This is what I think the market still underestimates. Part of the next capability jump may come not just from training bigger base models, but from giving reasoning models more inference-time budget and better ways to use it. And once agents become useful, companies are not going to deploy them sparingly. They are going to push them into more workflows.

Not only does training want more compute but inference does too and both are scaling rapidly.

The Shortage Is Already Showing Up

This is not a theoretical story anymore. The shortage is already showing up in how the leading labs talk, operate, and raise capital.

OpenAI has been explicit that compute scarcity is forcing tradeoffs. Anthropic’s behavior around Claude Code is another clean real-world example. Once users start handing off real multi-step tasks to agents, the flat subscription model starts to break because the infrastructure burden is much heavier than normal chatbot usage.

The strain is showing up not just in model-training budgets, but in actual product design and billing decisions. When agents stop being a demo and start becoming a workload. It gets expensive, fast.

The numbers tell the same story. Sarah Friar noted in January that OpenAI’s compute grew from 0.2 GW in 2023 to 0.6 GW in 2024 to roughly 1.9 GW in 2025. Over the same period, annualized revenue went from $2 billion to $6 billion to more than $20 billion. Her line that revenue directly tracks available compute is one of the most important data points in the whole sector. And the trajectory has only steepened since then.

OpenAI just closed a $122 billion funding round at an $852 billion valuation and is now generating roughly $2 billion in revenue per month. In its own announcement, the company called durable access to compute “the strategic advantage that compounds across the entire system.”

Compute infrastructure is the whole game.

The Buildout

This is another area where I think the market still underestimates the scale. The combination of better-than-predicted pre-training scaling laws and the increase in agentic inference demand is what has triggered this historic wave of global capital deployment. Capex by the top five hyperscalers alone is now projected at roughly $750 billion in 2026, up around 67% year over year and the third consecutive year of growth exceeding 60%. CreditSights estimates roughly 75% of that spend is going directly to AI infrastructure. Morgan Stanley has estimated roughly $2.9 trillion in global data-center construction cost through 2028.

There is no question the demand is there. The announced pipeline is enormous. But the amount actually under construction is much smaller, and there is a massive bucket where the completion date is still unknown. That is the tell. The buildout is happening, but a lot of projects are running into walls. Power. Permitting. Equipment. Labor. Time. The demand is clearly there. The hard part is getting the factories built. And capitalism tends to reward the people who solve hard problems with real demand on the other side.

Stargate is probably the single most visible example. The $500 billion joint venture between OpenAI, SoftBank, Oracle, and MGX is designed to deliver gigawatt-scale AI compute infrastructure across the United States, with the flagship Abilene campus already partially operational and sites under development in Wisconsin, Michigan, Ohio, and elsewhere. It is also, not coincidentally, running into exactly the bottlenecks you would expect at this scale.

Financing has been complicated. Power delivery timelines are long. A winter cooling failure at Abilene damaged relations between OpenAI and its development partner Crusoe, and OpenAI ultimately dropped plans to expand the flagship site further in favor of spreading capacity across other locations.

Stargate is both the strongest proof point for the AI factory thesis and a real-time illustration of how hard it is to get them online.

Addressing the Counterarguments

There are two common pushbacks to this view that are worth addressing directly.

The first is efficiency. A lot of people hear about better inference efficiency and jump to the same conclusion: less hardware needed, less infrastructure needed, thesis broken. I think that is backwards.

Take Google’s recent TurboQuant research paper. The core message is that long-context inference can get much more efficient. Fine. But if long-context inference gets dramatically cheaper, companies aren’t going to say “great, let’s use less AI.” They are going to say “great, now we can actually afford to use it for more things.” Bigger context windows become practical. Persistent memory becomes practical. Tool use becomes practical. Background agents become practical. More reasoning per task becomes practical.

So yes, efficiency improves. But that does not automatically mean demand falls. It can just as easily mean demand expands: Jevons Paradox.

The second is overcapacity. Every infrastructure boom sounds smart until it doesn’t. Dark fiber in 2001 is the usual comparison. But I do not think that is the right analogy. The better way to think about the current cycle is that demand is arriving before supply, not the other way around. A large share of current capacity under construction is already spoken for. That is what the pre-leasing data keeps showing. Microsoft’s Azure commentary points in the same direction. The company has now said multiple quarters in a row that demand exceeds available supply, even as more capacity comes online.

That does not mean every project gets built on time or every dollar of capex earns a good return. It does mean this is not a case where capacity is being built into a vacuum. The demand is already there. The bottleneck is delivery.

What Could Make This Wrong?

There are obviously real risks to this view. Maybe the Mythos story is overstated. Maybe Grok 5 disappoints. Maybe coherent compute matters less than I think. Maybe inference efficiency outruns deployment demand. Maybe enterprises move slower than the bulls expect. Maybe parts of the current buildout get ahead of near-term monetization.

All fair. But I still think the burden of proof is shifting, because the evidence coming out of 2026 is not pointing toward a world where scaling is dying and infrastructure demand is quietly fading. It is pointing toward a world where the frontier may still be improving faster than expected, while useful AI on the deployment side is becoming much more compute-heavy.

Final Thoughts

I think the market is still anchored to the wrong mental model. It still wants AI to behave like software. More and more, AI is starting to behave like an industrial business amassing a fleet of factories.

That is why I think the most important question in AI right now is what happens if training scaling laws are better than forecasted, the next step change requires massive coherent compute, and agents make inference much more compute intensive?

In that world, there will be plentiful business for those who can actually build the AI factories. And the winners will be the ones who can navigate through the full stack of problems: power, land, permitting, construction, cooling, and timing, all at once.

Leave a comment

Subscribe now

Share

Executive Summary

• Labor is the next real bottleneck in AI data centers. Hiring is already strained, delays are rising, and even serious operators like DataBank are saying peak labor needs have exploded far beyond what the current skilled-trades pool can support.

• Hyperscalers are feeling the squeeze too, they are just rich enough to hide it better. They can pay up, pull labor from farther away, and fund their own workforce pipelines, but even they are increasingly brute-forcing around a structurally tight labor market.

• Project ATOM matters because it directly addresses the labor bottleneck. If NUAI can cut field labor by roughly 80% and compress time to delivered powered shell, it can materially improve both competitiveness and economics.

• The shell is the right layer to standardize. Compute changes too fast, but the shell, structural carrier, and MEP backbone change much more slowly, which makes them much better candidates for repeatable industrialization.

• The real bet is not modularity by itself, but a platform that removes bottlenecks. If NUAI can combine land near existing power infrastructure, BTM power, modular shells, financing, and repeatable deployment into one integrated system, then it starts to look much more like a scalable platform than a one-off project.

The Second Bottleneck

For a while, the AI boom was a GPU story.

Then it became a power story.

I increasingly think the next bottleneck is human.

The capital is there. The demand is there. What is not there, at least not in enough quantity, is the specialized labor required to physically build these campuses.

Most investors still do not fully appreciate that.

Bill Stein put it plainly on the Data Center Revolution podcast:

“First, they talk about the shortage of power. I think the shortage of labor is the biggest problem we have in spades.”

“I do think that labor is going to be the biggest bottleneck. If you look out over 10 years, I think we’re going to always continue to wake up where skilled labor is our biggest challenge.”

“There’s just not enough talent for what the requirements are.”

The hard data points in the same direction. The 2025 AGC/NCCER workforce survey found that 92% of construction firms were having trouble filling open roles, 45% said labor shortages were the leading cause of delay, and 78% reported at least one delayed project in the prior year. ABC said in January 2026 that the construction industry would still need to attract 349,000 net new workers in 2026 and 456,000 in 2027 just to keep supply and demand in balance.

Where exactly are those people supposed to come from?

That question gets sharper when you look at data centers specifically. Tony Qorri at DataBank said sites that once peaked around 750 workers are now hitting 4,000 to 5,000 workers, and that the industry simply does not have enough qualified workers to meet demand. DataBank also warned that labor shortages, along with supply bottlenecks, are creating compounding risks in 2026. Oracle’s OpenAI-related projects were reportedly pushed from 2027 to 2028 largely because of labor and material shortages.

The hyperscalers are feeling it too. They just have balance sheets large enough to hide the problem better than everyone else. Microsoft has said it hired thousands of electricians across key data center states, with some workers commuting long distances or temporarily relocating. Google is training skilled trades through STAR. AWS is funding fiber, operations, and technical-trades pipelines in states where it is building. In other words, the biggest players are not immune. They are brute-forcing around the constraint by throwing money at the problem.

And I do not think this chokepoint gets looser from here. More AI infrastructure is likely to stay concentrated in the United States. The #1 economy in the world can raise capital much faster than it can produce electricians, welders, pipefitters, and controls specialists.

Furthermore, U.S. construction already relies heavily on immigrant labor, which means tighter immigration enforcement only makes an already constrained labor pool tighter.

At some point, money stops being a “good enough” solution.

That is why ATOM matters.

It does not require labor scarcity to become apocalyptic. It only requires labor scarcity to remain structurally tight enough that reducing field labor, standardizing more of the build, and moving more work into factories remains economically superior.

If NUAI solves both the power and labor bottlenecks, the story starts to look much bigger than one site in West Texas.

It starts to look a lot more like a platform built around a single north star: delivering powered shells faster than anyone else.

The Core Idea

If NUAI can standardize a repeatable 25 MW AI-ready shell, move a meaningful share of assembly offsite, pair it with powered land and a flexible BTM power strategy, and finance deployments through its GP/LP structure against long-term contracted hyperscaler cash flows, then it may be able to compress time to delivered powered shell in a way that materially moves the needle economically.

It matters for three reasons.

First, it helps NUAI win deals. The highest priority for hyperscalers is time to online compute, assuming reliability is a given. The developer that can deliver a faster powered shell should have an advantage both in the pitch and at the negotiating table.

Second, it pulls cash flows forward. Earlier delivery means earlier rent. In a DCF where the cash flows are still years out, that can materially increase NPV and present fair value per share.

Third, it may allow NUAI to retain more of the project-level equity. If modularization lowers shell CapEx, reduces field labor dependence, and compresses the construction schedule, total capital required comes down. That should improve project IRRs and may allow more of the equity value to flow back to NUAI, both because less equity may be required upfront and because developer promote hurdles become easier to clear.

What Exactly Is Modular?

The phrase modular data center gets thrown around too loosely.

Historically, most of the modular conversation in data centers was really about modular compute: containerized server blocks, packaged IT modules, edge deployments, and transportable units that could be dropped into the field quickly.

That is not what Project ATOM is building.

ATOM is a modular powered shell.

In other words, not just standardized walls and roofs, but a repeatable enclosure plus a large share of the core MEP architecture, cooling systems, and electrical backbone. Pair that with powered land and the result starts to look like a standardized powered shell delivery platform.

That does not mean every layer is frozen.

Quite the opposite.

Compute changes quickly. Rack power changes quickly. Cooling interfaces change quickly. Network architecture changes quickly too. The point is not to standardize every component forever. The point is to standardize the right layers: the structural carrier, the shell, the major distribution paths, and as much of the powered shell architecture as possible while keeping the tenant-specific equipment stack flexible enough to evolve with each generation of AI chips, rack density, and networking.

Put simply, the answer is not to standardize the compute.

It is to standardize the carrier around the compute.

First Principles

From first principles, the case for modular powered shells is fairly simple.

Traditional data center construction is still too close to bespoke field assembly for a product the market increasingly wants in repeatable blocks. You are taking a mission-critical industrial asset, building it outdoors, sequencing dozens of trades on a live site, and exposing the job to weather, site logistics, labor coordination, rework, and inspection bottlenecks. No wonder timelines slip and costs drift.

That model made more sense when demand was lower, designs were more bespoke, and speed mattered less.

That is not the world we are in anymore.

If the powered shell is becoming more repeatable, then the build process should become more repeatable too.

There is real evidence this is already happening at the shell, MEP, and prefabricated infrastructure layers. Turner & Townsend’s 2025 data center cost index says 53% of respondents expect the majority of future high-density projects to include liquid cooling, and UL’s updated UL 2755:2025 now evaluates prefabricated modular data center systems as integrated products rather than one-off custom assemblies.

The manufacturing logic follows from that. Move more work into a factory, standardize more of the design, pre-test more of the system before it hits the site, buy materials in larger and more predictable batches, and reduce how many activities still have to happen sequentially in the field.

In other words, modularity changes the physics of the build.

Instead of waiting for the site to be fully ready before everything starts, you can parallelize. While land is being cleared, foundations are being poured, turbines are being set up, and utilities are being prepared, major portions of the shell and MEP can already be moving through a controlled manufacturing environment.

That should theoretically significantly compress the construction schedule.

It should also reduce variability, which is just as important. Factory work is not magic, but it is almost always more controlled than field work. You get repeatable workflows, repeatable QA, repeatable tooling, better supervision, less weather disruption, and fewer chances for one trade slipping to cascade into six others.

This is also not just a quantity problem.

It is a skill-mix problem.

AI campuses are not simple boxes. They are increasingly dense, increasingly liquid-cooled, and increasingly dependent on tighter tolerances and more specialized installation work. You can build more assembly lines and expand fabrication capacity. You cannot instantly create thousands of experienced electricians, welders, controls specialists, and pipefitters.

That is why modular powered shells make so much sense. They are, in part, an attempt to use manufacturing to get around the fact that the field labor pool cannot scale fast enough.

If the product repeats, the process should repeat.

And if the process repeats, it should get faster, cheaper, and more reliable over time.

A Brief History Of Modularity In Data Centers

The first real modular wave in data centers came from the early cloud giants, especially Microsoft and Google, around the late 2000s. The goal was obvious: industrialize capacity instead of building every facility like a one-off project. That wave mattered, but it standardized the wrong layer. Container-style modularity lived too far downstream, at the compute layer itself, exactly where the technology was evolving fastest, and the box itself became the constraint as power density, cooling requirements, and networking all changed.

That is the real lesson from the first modular cycle.

Not that modularity failed.

That the industry standardized the wrong layer.

The good part of the original vision survived. The market kept the standardization, repeatability, prefabrication, and parallelization, then pushed those ideas higher up the stack. Over time, the industry moved away from shipping containers full of compute and more toward prefab electrical rooms, cooling skids, white-space pods, and increasingly standardized shell and MEP systems.

BladeRoom, a U.K.-based provider of modular, shell-centric data center systems built around stackable blocks of up to 30 MW each, is the clearest bridge between the old modular world and the current one. It proves shell-centric modularity is real. The category is not theoretical.

The field is not empty either. Old modular compute proved factory-built delivery can work. Prefab incumbents prove modularity is already mainstream in parts of the stack. What makes RK and ATOM more interesting is the attempt to pair a modular shell concept with real manufacturing and system integration.

Project ATOM

Based on NUAI’s March 2026 presentation and Q4 2025 earnings call, ATOM is being framed as a repeatable 25 MW base deployment unit designed to scale from a single block all the way to multi-gigawatt campuses.

More importantly, it is being pitched not just as a modular building design, but as a standardized product platform.

That distinction matters.

A lot of data center designs are still effectively one-off projects with a few standardized components layered in. ATOM is trying to go further than that. The pitch is one proven base unit repeated over and over, with enough built-in flexibility and buffer margin to absorb future hardware upgrades and density increases without breaking the platform.

The most differentiated disclosed features are:

• 25 MW base deployment unit

• 80% modular construction

• 45 to 135 kW per rack

• a two-story architecture with MEP systems stacked above the IT load

• a claimed 15% to 20% lower energy draw from shorter piping runs

• 0.15 WUE and 90%+ free-cooling hours

Those are impressive specs.

But individually, they are not a moat.

What matters is the combined system: a repeatable 25 MW shell, heavy modularization, factory-first construction, a standardized architecture, and a design built explicitly for high-density AI.

The timing matters too. NUAI says manufacturing is planned to begin in 2026, with first sites expected to be energized in 2027 or 2028.

ATOM is also not just trying to standardize the physical shell. The company is describing a standardized operating layer around it: engineering reuse, a common BMS platform, a common vendor stack, and shared deployment processes.

That is why the platform framing is more interesting than the spec sheet.

The upside is not one building built a little faster.

It is a system that should iterate and get better each time it is repeated.

Why RK Mission Critical?

Did NUAI choose the right manufacturing partner to turn ATOM from concept into a repeatable product?

Publicly, RK Mission Critical looks much more like an industrial integrator than a catalog modular vendor. The company says it has deployed more than 2 GW nationwide, operates roughly 402,000 to 422,000 square feet of manufacturing space depending on the source, and runs 11 production lanes. It also emphasizes factory QA, pre-testing, and repeatable workflows.

The challenge with modular data center shells is not renderings. It is building them repeatedly, with high quality, speed, and enough throughput for the unit economics to improve over time.

The partnership also appears to offer more than generic access to a contract manufacturer. NUAI’s materials say the relationship provides qualified production capacity in controlled manufacturing facilities, helping ensure consistent and timely supply.

In a modular model, the bottleneck can move from the job site to the factory.

So qualified access to capacity matters.

RK also has at least some evidence of actual modular mission-critical delivery, not just theory. Public materials reference a campus-scale project involving 70 purpose-built modules delivering 32,000 square feet of white space plus supporting mechanical and electrical infrastructure. More notably, RK highlights a customer testimonial saying initial energization was achieved just 47 days after breaking ground.

I would not build an entire thesis around one testimonial.

But I also would not ignore it.

In a story where time to power is everything, that is exactly the kind of datapoint you want to see.

More broadly, RK looks aligned with what ATOM is actually trying to become. A lot of modular vendors in this market are really shipping discrete IT, power, or cooling modules. RK looks different because the differentiation appears to be system-level integration: structure, mechanical, electrical, manufacturing, testing, and site coordination under one roof.

That is the whole point. The value is not in any one component. The value is in integrating the full powered shell stack tightly enough that the factory actually substitutes for field labor instead of just moving the complexity somewhere else.

What Has To Be True - And What Could Break It

One place I would be careful not to overstate the case is IP.

Based on what is publicly disclosed, I do not think the strongest moat argument here is pure shell IP by itself. Public materials do not clearly spell out the exact ownership split between NUAI’s platform IP and RK’s manufacturing IP, nor do they disclose broad exclusivity terms that would clearly prevent adjacent competition.

That does not kill the thesis.

It just means the moat needs to be framed correctly.

To me, the stronger moat argument is a platform moat, not a patent moat. Very few players can combine powered land, a credible power strategy, the right manufacturing partner, repeatable shell design, financing structure, and real customer access at the right time. The moat, if it emerges, is more likely to live in system integration, timing, and execution rather than in one magical patent.

For ATOM to matter economically, a few things have to be true.

• It has to be meaningfully faster than already-good conventional delivery, not just faster than average developers.

• The savings have to show up at the whole powered shell level, not just in one narrow slice of the build.

• Repeat volume has to be high enough for the factory curve to kick in, and that only works if the 25 MW block maps cleanly to how customers actually want to phase capacity. If each deployment has to be re-designed, the economics unravel quickly.

• Long-lead equipment procurement has to be controlled well enough that faster shell construction actually pulls forward energization and revenue.

Inverting the problem, the strongest bear argument is not that modularity is fake. It is that top developers may already be capturing most of the easy gains through conventional shells plus aggressive prefab MEP, without taking on the extra transport, lifting, code, and interface complexity that comes with more modularized building systems.

The problem is that ATOM is not competing against slow amateurs.

It is competing against increasingly optimized conventional delivery.

The bear case gets stronger when you look at where schedules actually break. Faster shell construction does not matter much if switchgear, generators, transformers, turbines, and cooling equipment still dominate the critical path. And the historical warning from the first modular wave still applies: standardize the wrong layer, or force too much redesign back into the system, and the economics can drift right back toward traditional construction.

That is what would break the thesis.

If early deployments do not show real delivery advantages, if redesign kills reuse, if long-lead gear still controls the schedule, if hyperscalers prefer other delivery models, or if NUAI cannot win real deals that turn ATOM from concept into repeatable deployment, then the thesis gets materially weaker.

Conclusion

Before ATOM, NUAI looked like a potentially valuable powered land and time-to-power story.

That alone is interesting.

But after diving deep into Project ATOM and modular data centers more broadly, I think the upside may be bigger than that.

Why?

Because power is only the first bottleneck.

If labor is the next real bottleneck, and I increasingly think it is, then the value shifts toward whoever can remove execution bottlenecks in a repeatable way.

That is what makes ATOM interesting.

NUAI first attacked the power bottleneck.

Now it is trying to attack the labor and construction bottleneck too.

If it can do both, then this starts to look like something much bigger than a one-off West Texas project.

It starts to look like a platform.

More specifically, it starts to look like NUAI is building the delivery operating system for powered shells.

That is the real product. Not just the shell itself, but the machine behind it: site selection, powered land, BTM strategy, manufacturing partners, financing structure, design standards, and deployment processes integrated into one system that removes bottlenecks one by one.

That is a much bigger idea than modular construction by itself.

A powered land story can be valuable.

But a repeatable system for delivering powered shells at speed can be worth much more.

Because if ATOM works, NUAI is no longer just a company with one good site.

It is productizing a solution.

And that solution is not just faster power.

It is faster power, faster shell delivery, lower dependence on constrained field labor, more repeatable execution, and a platform that should get better with each iteration.

That is where the real upside is: turning what has historically been bespoke project work into something much closer to a repeatable deployment machine for hyperscalers.

And if that happens, the moat will not come from one piece of IP.

It will come from the full stack.

Speed to power.

Speed to shell.

Financeability.

Design reuse.

A repeatable deployment process.

And the ability to turn hard-to-build energy corridor sites into institutional-grade AI campuses faster and more predictably than conventional alternatives.

Put simply, if ATOM works, NUAI is not just building a better box.

It is building the machine that removes bottlenecks from rapid powered shell delivery.

Disclosure: I am long NUAI. I may buy, sell, or change my position at any time without notice. This article reflects my personal opinions, is for informational and educational purposes only, and is not investment advice or a recommendation to buy or sell any security. I have not received compensation from NUAI, its management, or any affiliated party for this content. I have no business relationship with the company.

View my full NUAI model and price target here.

View my deep dive on NUAI’s Q4 2025 earnings call here.

Share

Subscribe now

“This is not just an earnings recap. It’s in our filings. The numbers are public, and we can all read them. What this call is about is something more fundamental. We’re going to walk through what New Era Energy and Digital actually is, how it’s built, and why we believe we’re positioned to be a defining platform in AI infrastructure development.”

Intro

First, I want to apologize for my late commentary on the Q4 management call that occurred on March 17th. I have been backpacking without access to the internet.

That being said, despite the limited time and connectivity, I wanted to get some of my thoughts out to readers.

Like any other shareholder, I would have loved to see a hyperscaler FID during the call. To be entirely honest, I had expected management to abide by their self-imposed Q1 2026 deadline.

Alas, I should have probably expected a delay given management’s history of pushing deadlines (e.g. Will Gray already moved the FID deadline from Q4 2025 to Q1 2026).

Management has now set a pattern of overpromising and underdelivering. I don’t like that, but it reminds me, however imperfectly, of Elon’s style: aggressive deadlines, frequent misses, and relentless iteration. The lesson is not that missed timelines are good. It is that timing matters less than outcomes. And despite repeatedly missing deadlines, Elon’s companies have still produced some of the most impressive hard-tech innovations of the modern era.

Fundamental business advancement drives value creation which, over time, dictates the price the market is willing to pay for said business.

While there is still time left in Q1, I have entirely reset my expectations for the FID announcement due to management’s commentary during the call. Will made it clear that it was a mistake to have announced a FID deadline in the first place as the exact timing is more than likely out of NUAI’s control.

Management made it clear that talks were materially advancing and going well, but that massive bureaucracies move slower than a lean startup team like NUAI.

What actually drives value?

In other words, what actually drives the model? The exact day/month when the FID comes out does not dramatically change the inherent value of TCDC and thus NUAI. While that sounds crazy and pretty much every shareholder is banking on an FID to re-rate the stock, it does little in the grand scheme of the broader TCDC project valuation if you believe these things to be true (which I do):

1. If you believe management is broadly being truthful about the status of negotiations and the timeline to first power, then the valuation framework still holds.

2. You believe the re-iteration made by management on the call that the time to first power date (Q3 2027) is unchanged.

Regarding bullet one: many, including some shareholders, believe that CEO Will Gray is untrustworthy. Given his past business turbulence as a Landman, the New Mexico legal overhang, and the repeated FID deadlines missed I honestly can sympathize.

What I simply cannot understand is why, if NUAI were truly overstating the seriousness of its negotiations and was not actually in advanced talks with a Big 4 hyperscaler tenant, someone like Ted Warner, the new CFO, would choose to join a company like this.

Why would a highly successful banker leave a rainmaker role, likely earning seven figures, after working with companies like APLD and helping close over $7 billion of HPC financings, to join a cash-poor company on a deal where most of his compensation is performance-based equity tied to that company actually landing a hyperscaler FID and monetizing 250 MW of compute?

It simply does not make sense.

During the call Ted himself said that:

“I’ve seen hundreds of opportunities in HPC and when I dug in here I knew that I found exactly what I was seeking.”

Get that, the man who massively aided in funding the highly successful powered land & shell company that NUAI is trying to emulate in terms of success decided that NUAI had a good enough shot to become a scaled competitor that he used his single most powerful vote of confidence, his career, on NUAI.

Ted Warner’s decision to join makes the bear case that NUAI is materially overstating the seriousness of these negotiations much harder to believe.

Furthermore, for what it is worth, I believe that Ted is exactly who NUAI was missing: someone who can help raise the much needed capital at both the hold-co and project level. He raised the billions in capital that APLD needed when they were a recent crypto pivot small cap with lofty HPC aspirations. I believe that he can do the same for NUAI.

Enough with that tangential sidebar. Point being I like the new CFO.

Again, what drives the model, and therefore intrinsic value, is when compute actually comes online and begins generating cash flows through the NNN lease. The sooner those cash flows arrive, the more they are worth to present shareholders. If time to power gets pushed out, NPV declines because of the time value of money.

That is what drives intrinsic value, not a press release naming the hyperscaler. As of now, time to power is unchanged, and therefore my thesis and valuation remain unchanged.

One worry shareholders may have, however, is that the hyperscaler, without a public binding agreement, will simply back out leaving NUAI without a tenant.

I believe this to be a very real risk but highly illogical. Why would a Big 4 hyperscaler all of whom have repeatedly cited time to power as their #1 priority, with record amounts of CapEx, and massive cloud backlogs back out of one of the quickest time to power projects in the 450MW+ range?

The demand for AI compute is simply too high.

I believe CEO Will Gray, particularly his Q&A comments where he explicitly said the FID is progressing and will happen.

“We are pretty far along… Is it progressing? Absolutely. Will it happen? It will happen. And I do feel like there’s going to be some news that we can announce forthcoming that to give some clarity.”

Something else that caught my eye was a commitment to begin construction in Q2 2026. This is great news and likely suggests a very advanced LOI / working agreement or an imminent FID.

Again, what I care about as a shareholder is when TCDC will start generating cash flow. Construction, to me, matters far more than any FID or PR announcement.

Actions are much louder than words.

I look forward to seeing, either in person or via photos, the progress made at the TCDC site in the coming months.

NUAI has assembled all of the pieces of the puzzle; it is time to execute.

Subscribe now

The Infrastructure PE Model

As we have discussed here, and as individuals such as Dulce on X have also covered, the GP/LP model is a clear differentiator for NUAI. I was pleasantly surprised to see management cover it in the call in such explicit detail.

If you are not familiar, the GP/LP model basically allows NUAI (as the GP) to earn carry (or sweat equity) in projects via their contributions such as the land and actually developing the project via subcontractors. Because of this they are able to, like private equity firms in infrastructure PE, to walk away with significant equity stakes in these projects without having to put up very much, if any, upfront cash.

This exact dynamic is what allows PE firms that are themselves worth maybe only tens of millions to do deals where they hold significant equity stakes in businesses worth hundreds of millions, if not billions, of dollars in enterprise value. More importantly, they are not just able to do these deals once. They are able to scale and pursue as many projects as they can source, often limited more by manpower than capital, assuming they have sufficiently accretive deals to offer investors (LPs).

In other words, while companies such as IREN and APLD are forced to raise billions of dollars through dilution and can only work on a handful of deals at a time because their balance sheets can only stretch so far, NUAI, through the GP/LP model, could theoretically pursue orders of magnitude more deals simultaneously.

Management emphasized how this exact model funded the midstream (e.g. pipelines) buildout during the shale boom from 2010-2014.

Billions of infrastructure built with the balance sheets of pension funds, endowments, sovereign wealth funds, and family offices. Billions of fees and equity collected by the developers.

“Those infrastructure companies didn’t produce a single barrel of oil, and they didn’t consume gas. What they did was build the infrastructure that connected the supply to the demand… It was financed through a very specific engine: project-level financing, GP/LP structures, and asset-level debt secured against long-term contracted cash flow… This is what NUAI is doing: applying that exact playbook.”

“So essentially, we are the midstream of the AI economy.”

The closest historical parallel is probably early Kinder Morgan and the broader MLP-era midstream buildout. The similarity is not just that both businesses sit in a bottleneck layer of infrastructure. It is that both rely on the same core financial engine: outside LP capital and asset-level debt funding the project, while the GP keeps the upside through fees, promotes, and retained equity ownership. Kinder did not need to fund every mile of pipe off its own balance sheet to create enormous value. It needed to control the project, structure the capital stack correctly, and lock in long-duration contracted cash flows.

That is a very similar playbook to what NUAI is trying to do just for the AI boom.

NUAI’s Differentiated Land Model

Another bit that I found interesting was management’s description of how they are differentiated in their sourcing criteria for land.

While traditional REIT developers have historically sought out only the most pristine plots of land, not just in terms of PPAs and grid connectivity, but also in the actual characteristics of the land itself.

For example, issues such as existing wells and mineral rights that need to be renegotiated, abandoned piping and other legacy materials, or a lack of existing fiber infrastructure would have been an instantaneous pass for these firms.

NUAI brings a “fixer-upper” operator mindset to land development that feels borrowed straight from the oil and gas industry. This “hair” on the deal, the very thing that caused traditional developers to pass, is exactly where NUAI appears to derive value. As with TCDC, NUAI does not seem afraid to put in extra sweat equity by clearing out old pipes, renegotiating mineral rights, remediating old wells, and so on.

Management described this on the call as turning C sites into A+ sites. That may sound like table stakes, but in a world where many of the premier data-center-ready plots in the U.S. are already being developed and/or bid up to astronomical prices, it gives me much more confidence in NUAI’s ability to scale into a truly repeatable platform if it can consistently source land that the vast majority of competitors have already passed on.

“We chose TCDC, which was riddled with pipeline easements, active pipelines, all sorts of things. And those to a normal developer would look undevelopable. But our group… it’s allowed us to basically play moneyball with sites that otherwise are very well positioned but hard developed…”

Furthermore, when you think about the capital-light GP/LP model, the source of NUAI’s initial equity stake is more than likely the appraised value of the land it brings to the table.

If NUAI is able to purchase a plot of land for, say, $100 million that others passed on because of the “hair” on the deal, then put another $25 million and a few months of work into cleaning it up and turning it from a C into an A+ site, the value creation can be significant. Once that land is reappraised in a state where the REITs typically fish, with massive amounts of capital behind them, it is not hard to imagine a 2x, 3x, or even 4x re-rating in the land’s value.

Take that to the negotiating table, and NUAI could walk away with an equity stake that far outstrips the initial $125 million invested to acquire and improve the land.

While legacy developers bid pristine plots up to sky-high valuations, NUAI is trying to manufacture wins by fishing in a blue ocean.

Project ATOM

To me, this was by far the most significant and exciting part of the call. Project ATOM appears to be Charlie Nelson’s brainchild, and I loved almost everything he described.

Again, the north star for NUAI is delivering the fastest time to power possible.

We have already seen NUAI address the grid interconnect backlog through BTM power, truly being one of the first companies to focus on this solution.

Now, in trying to pierce through the next bottleneck, NUAI is moving on to solve the labor and construction side of the data center buildout process through Project ATOM.

The vision, and one I fully believe is where the industry is headed, is a data center model where 80% of the buildout is completed through a highly efficient, repeatable, and reliable factory process. Each module would be built in a kind of assembly-line format in 25 MW blocks.

“What we’ve identified as a very critical weakness industry-wide is… these large data centers take an enormous amount of labor… you can only build so many of these at the same time before you stress the labor pool until it breaks… So what’s the answer to that? The answer to this is standardization and modularization.”

“The Atom platform is a 25-megawatt-based unit… 25 megawatt units can be stacked side by side… And what it allows us to do is reduce our on-site labor by about 80%, as it’s being projected today. But it also means that most of this is constructed in a controlled manufacturing environment…”

Now, not only do you enable the automation of much of the buildout process, thereby addressing one of the biggest bottlenecks facing nearly every new data center project today, namely labor shortages, but you also benefit from economies of scale because materials can be purchased in massive bulk orders. On top of that, you can continue iterating on the manufacturing process, which is highly standardized to pre-approved hyperscaler specifications aside from a few tweaks as next-generation chips are released, until you are not only building faster, but also cheaper.

To execute Project ATOM, NUAI has signed an LOI with RK Mission Critical, a part of the much larger RK Industries, which management described as “a world-class modular data center manufacturer capable of completing approximately eighty percent of assembly within controlled factory environments.”

My initial read is that RK Mission Critical will be able to deliver on what NUAI is envisioning with Project ATOM. They appear to have been working on modular data centers for more than five years and seem to have real-world execution experience, albeit at smaller scale.

Publicly, RK Mission Critical says it has already delivered over 20 indoor modular Tier III units for a world-class data center services provider, and one recent campus-scale project included 70 purpose-built modules delivering 32,000 square feet of white space plus the supporting mechanical and electrical infrastructure. On top of that, RK says it operates 422,000 square feet of manufacturing space across 11 production lanes.

Charlie also stated that RK already has dedicated production capacity set aside for NUAI and has been very clear about what it can produce.

Interestingly, the broader RK platform has real hyperscaler adjacency, including publicly disclosed work on Microsoft’s CY01 data center.

A few other important points: RK has over 2,000 employees and is vertically integrated all the way down to steel manufacturing. Furthermore, NUAI worked with EYP/Ramboll, the same engineers involved with TCDC, to custom design the ATOM platform.

In Charlie’s words, this modularity enables the “copy and paste manufacturing” of AI data centers.

To be clear, modular data centers are not a new concept, and NUAI is not unique simply because it is pursuing modularity. Where I believe the model starts to look differentiated is in the packaging: full powered-shell style deployment, custom hyperscaler-oriented 25 MW building blocks, dedicated manufacturing capacity through RK, and all of it layered on top of a power-first GP/LP development model. In other words, the novelty is less the idea of modularity itself and more the specific way NUAI is trying to industrialize and finance it.

ATOM allows NUAI to parallelize work in two ways: first by building the shell while the land is still being cleared, and second by manufacturing multiple modular units at the same time across assembly lines.

Much like GPUs parallelize compute and execute tasks far faster than CPUs, which solve them sequentially, NUAI is trying to parallelize data center development.

Liquidity

Liquidity is still the near-term overhang. Management ended the year with just $1.2 million of cash, and even after giving credit for post-year-end warrant proceeds, my rough estimate is that NUAI probably has only around $5–$5.5 million of cash today after burn.

Simply put, that is not enough to cover the $9.85 million cash payment due to SharonAI by March 31, alongside the separate $10 million equity obligation.

Will Gray addressed this directly in the Q&A, pointing to a “senior convert,” “mezz piece,” or “credit facility” as the likely short-term funding path, while also acknowledging the month-end SharonAI obligation.

I expect NUAI to plug this gap with some combination of bridge capital and equity-linked financing. That would likely involve some dilution, but in the context of a project this large, I would view modest dilution here as a financing bridge.

The overall thesis still holds.

Additional Takeaways

There were also a few other items from the call worth highlighting.

First, Ted explicitly cited an 80% LTV at the project level for TCDC. That is more aggressive than the 65% I had been modeling and, if achievable, would imply less equity capital required at the project level and more attractive sponsor economics.

Ted also laid out the monetization stack clearly: development fees upon project delivery, recurring management fees for ongoing asset management, and the right to material equity participation in each project.

My current model only assigns value to the last piece.

I have not yet included development or management fees because I still need more color on the likely range, but for now that leaves another layer of conservatism in the valuation.

Second, there does appear to be real optionality beyond TCDC, even if I am not underwriting it yet.

Management said the New Mexico site is already undergoing engineering work by EYP/Ramboll and that NUAI is in talks with hyperscalers about potentially being a tenant. They also referenced a partnership with Last Energy tied to 1+ GW of SMR development.

I am still skeptical of this site given the legal overhang in New Mexico and the gargantuan 7 GW headline power figure, which is why my model continues to assign it zero value today. Still, it is clearly something worth watching.

More broadly, management also said it has a number of other assets ready to be spun up, while making clear that TCDC remains the absolute priority until it is locked down.

That is not enough to underwrite, but it does incrementally increase my confidence that NUAI may be building a platform rather than a one-off project.

Finally, a few outside signals also stood out to me.

I recently came across an industry research note from Jefferies pointing out that as other behind-the-meter projects run into obstacles, both second-hand and OEM gas-turbine supply should improve over the next several years.

That is not especially relevant to TCDC itself, since NUAI already has its turbines locked down through Thunderhead, but it is quite relevant to the repeatability of the broader model. I had been very concerned about how NUAI would source generation equipment for future projects, and this materially eases that worry.

Conclusion

To be completely honest, I was initially very disappointed with the call. Not because anything bad came out, but rather I had psyched myself up about a possible FID.

But this is why writing out my thoughts in these deep dives is immensely valuable for me.

As I continued to comb through my notes from the call and really dug into Project ATOM, I am coming away far more bullish than before.

Possibly even more bullish than if we had gotten an FID on that call.

I would summarize my thoughts like this: NUAI becoming a successful platform is worth at least an order of magnitude more than a one-off TCDC success. I am materially more excited and confident that such an outcome is actually possible.

I would say before this call my personal confidence in NUAI actually being able to become a platform, regardless of even a successful 1 GW TCDC, was maybe 10-15%.

I would say I now sit somewhere closer to 35-40%. This will only continue to grow as NUAI continues to execute.

“Everything we’re talking about here is a repeatable development platform. Standardized frameworks to get faster and better with every single project we do. TCDC is project one, but the platform that comes with the partnerships, the playbook, financing model, the permits, the execution, all of that is designed to support a lot of campuses over time…”

If management is actually able to execute on Project ATOM, NUAI could develop a true, durable, and wide moat with meaningful differentiation from competitors.

It would become very hard for me to sell a single share of such a company.

Until then, I will be patient. I suggest my fellow shareholders do the same.

To Will, Charlie, and Ted:

You have crystallized the vision. Now it is time to execute.

View my full NUAI model and price target here.

View my full earnings call transcript (raw & edited) here.

Leave a comment

Disclosure: I am long NUAI. I may buy, sell, or change my position at any time without notice. This article reflects my personal opinions and is for informational and educational purposes only. Nothing in this article should be construed as investment advice or a recommendation to buy or sell any security. I have not received compensation from NUAI, its management, or any affiliated party for this content. I have no business relationship with the company.

Intro

NUAI released its FY2025 10-K.

At a high level, I think this 10-K does three things.

1. It significantly weakens one of the main bear narratives around management misconduct and the New Mexico lawsuit filed against CEO Will Gray.

2. It aggressively resets the accounting around the legacy asset base.

3. It solidifies TCDC as the company’s core and absolute focus.

Subscribe now

Legal Update

In December, the stock price got hit hard when the State of New Mexico filed a civil action tied to old Acacia Resources wells, plugging liabilities, and affiliated parties including CEO Will Gray. Bears immediately weaponized that lawsuit to paint the whole company as toxic.

According to the 10-K, the independent members of the Board initiated an internal investigation with independent outside counsel. It says “No limits were placed on the scope of the investigation.” And most importantly, it says the investigation “found no facts supporting the allegations of wrongdoing.”

Here is the exact language:

“In response to the New Mexico Litigation and reports by purported short sellers on subject matters similar to those alleged in the lawsuit, the independent members of the Board promptly initiated and conducted an internal investigation into the allegations, with the assistance of independent outside counsel.”

“The investigation included review of documents and many interviews. No limits were placed on the scope of the investigation.”

“The investigation found no facts supporting the allegations of wrongdoing in the short seller reports or the New Mexico Litigation by Solis Partners, LLC, Mr. Gray, the Company, or any entities associated with Mr. Gray.”

That is a meaningful positive development for NUAI.

Obviously, it does not make the litigation disappear. The company is explicit that the matter was filed on December 23, 2025, was “stayed shortly after it was initiated,” remains in its “early stages,” and may still create legal costs or interfere with the sale of legacy assets. The filing says NUAI “may incur significant legal and other fees and costs,” is “not currently able to estimate the possible cost,” and that “we may be unable to consummate our contemplated sale of legacy natural gas assets” while the litigation is pending.

So no, this is not some clean all-clear.

But I do think it changes the burden of proof.

Bears can still point to the existence of the lawsuit. What is harder to argue now is that the company’s own internal review uncovered something obviously disqualifying.

The mechanics here were handled the right way. The investigation was initiated by the independent Board members and run with independent outside counsel. That does not guarantee a perfect process, obviously, but it is meaningfully different from management investigating itself and declaring a heavily biased victory.

The wording matters too.

Corporate lawyers usually hedge. They say they could not substantiate the claims at this time or that they did not identify evidence sufficient to support a conclusion. They do not usually say the investigation “found no facts supporting the allegations of wrongdoing” unless they are comfortable with that language.

My read is simple: the filing did not eliminate the legal risk, but it did materially weaken the strongest bear narrative.

The Legacy Assets

One thing this 10-K made clear was management’s aggressive reset of the legacy asset story and associated valuations.

The filing says:

“Impairment expenses increased $12,062,639 for the year ended December 31, 2025, as compared to the year ended December 31, 2024.”

And then breaks it down:

“The increase was primarily due to a $6,732,000 impairment of oil and gas properties as a result of a ceiling test failure, a $5,330,000 impairment of the gas processing plant.”

That is not subtle, and on a GAAP income statement, yes, it looks ugly.

But I think the more useful question is whether the company is finally forcing investors to stop carrying the legacy business at values management itself no longer seems to believe in.

And I think the answer to that is yes.

As for the Pecos Slope Plant, I believe, for all practical purposes, that this project is dead.

The 10-K does not quite use those words, but it comes close.

The filing says the plant was “unlikely to be completed or placed into service.”

And then comes the key line:

“The asset was written down to its estimated fair value, which was determined using a market-based assessment of expected recoverable value and was considered negligible.”

That is about as close to a death certificate as you are going to get in SEC language.

Again, I do not think this changes the strategic thesis.

I think it confirms what the strategic thesis already was.

In my downside liquidation case, I was already fully discounting the legacy assets to zero. This filing simply solidifies that decision and reiterates management’s full, all-in focus on TCDC.

The Reserves

The filing says the company recorded a ceiling test impairment of about $6.73 million and that it was driven by “a change in strategy” that resulted in the removal of the company’s proved undeveloped locations.

For reference, the reserve tables are ugly.

At year-end 2024, NUAI had 85,498 MMcfe of total proved reserves, including 55,622 MMcfe of proved undeveloped reserves. At year-end 2025, total proved reserves were down to 15,097 MMcfe, proved undeveloped reserves were zero, and the company had no probable oil and natural gas reserves.

The PUD reserve rollforward is even more telling. At December 31, 2024, the company had 80 proved undeveloped wells representing 55,622 MMcfe. By December 31, 2025, that number was zero.

And then comes the line that, in my view, really explains the whole thing:

“As of December 31, 2025, our two 10-year helium contracts had expired.”

That matters a lot.

The filing explicitly says the prior development schedule had been designed in part to support the Pecos Slope Gas Plant and in part to “fulfill the terms of our two 10-year helium contracts.” Once those contracts expired and management changed strategy, the reserve story collapsed with them.

Put simply, this was not a random accounting event.

It was the formal reset of a legacy story the market was already moving past.

Management is burning the boats on its legacy gas assets. Period.

What About Helium?

Given recent helium market tightness, it is tempting to look at NUAI’s 137,000-acre Pecos Slope Field and see a lottery ticket waiting to be cashed in.

I have seen a couple of takes on X arguing that NUAI’s helium assets have some needle-moving residual value due to the spike in helium prices driven by the closure of the Strait of Hormuz.

On paper, you can make the math work. If you assume roughly 1.5 to 2.5 Bcf of recoverable helium reserves and model out a production ramp with elevated domestic spot prices, you can pretty easily back into a project NPV north of $400 million.

It is a fun thought exercise.

But I think it is just a distraction.

Why?

Because a buyer is not going to pay NUAI for that theoretical upside. The buyer would have to take on the capital burden required to actually extract and process the gas. And on top of that, whoever buys these legacy assets also has to absorb the plugging and environmental cleanup obligations sitting on the balance sheet.

For reference, the filing shows asset retirement obligations of $12.319 million at year-end 2025.

The market can dream about helium spot prices all it wants, but distressed legacy gas assets do not usually get sold for blue-sky NPVs. They get sold to stop the cash bleed, shed liabilities, and let management move on.

Not even to mention the ongoing New Mexico lawsuit and management’s own admission that this may prevent them from exiting said legacy assets.

But yes, there may be macro helium optionality in some abstract sense.

I just do not think it is the right thing to focus on here.

It is just noise relative to TCDC.

The Signal

To me, the most important signal in the 10-K revolves around how management is describing the tenant stack, the power stack, and the economic model around TCDC.

Management is painting a fairly clear picture.

First, land an investment-grade hyperscaler anchor tenant to leverage their credit.

Then, once the early phases are stabilized and financed, layer in higher-yield tenants around that hyperscaler core.

The end result, at least in theory, is a fully leased 1 GW campus with a very efficient capital stack. You use the hyperscaler’s creditworthiness to support meaningfully more debt and lower the project’s WACC, while at the same time pushing your weighted average rental rate per MW upward by blending in riskier, higher-yielding tenants on the back end.

Here is the exact language:

“Our initial leasing efforts are rigorously focused on securing investment grade hyperscalers (e.g., major cloud providers and capitalization-weighted technology firms). We view these tenants as foundational to our capital strategy. Securing long-term, contracted cash flows from these counterparties is a prerequisite for stabilizing our asset base and accessing favorable debt financing terms for our project-level entities.”

“Following the stabilization of our initial phases with investment grade tenants, we intend to diversify our tenant base to include higher yield, higher-risk counterparties. This expanding cohort may include specialized graphics processing unit (“GPU”) cloud providers, sovereign AI clouds, and venture-backed AI startups. We believe that blending these higher-yielding tenants into a portfolio anchored by investment grade credit will allow us to optimize our weighted average return on invested capital while maintaining an acceptable risk profile.”

Management is explicitly telling you the first tenant is not just about occupancy. It is about financeability. Get the right name in the door, lock in long-term contracted cash flows, and use that to stabilize the asset base and unlock better debt terms. Only then do you start blending in the riskier, higher-yielding tenants.

Put simply, this is a classic anchor-and-expand model.

And if you are trying to read between the lines on who some of those higher-yield tenants could eventually be, the “GPU cloud providers” language stands out to me. My own view is that someone in the vein of Nebius would fit quite well.

The other subtle signal here is around which power source serves which tenant type.

“We plan to source a portion of our initial capacity from nearby merchant power facilities, which include the nearby Vistra and Quail Run plants, among others. Leveraging their proximity, we intend to construct direct grid interconnections that bypass broader transmission congestion, providing immediate baseload capacity.”

President and COO Charlie Nelson, on the BofA call, shot down the idea of using these plants directly to power GPUs because of their “jagged load.” But the 10-K language makes it sound like NUAI could still be tapping into this power in some capacity.

My read is that this likely points to a split campus architecture.

In other words, the nearby merchant or grid-connected power may be used to serve the more conventional, non-GPU, lower-density portions of the campus first, while the dedicated behind-the-meter generation is reserved for the actual AI workloads.

Use existing nearby baseload to accelerate early capacity delivery where possible, then match the most demanding racks to the dedicated on-site energy solution.

Again, that is my interpretation, not something management says outright. But I do not think the language is accidental.

And this ties directly into another important point in the filing: the energy silo is separate from the real estate silo.

“To secure dedicated behind-the-meter capacity, we are in advanced discussions with a leading provider to act as our primary on-site power partner. We intend to allocate a dedicated 20-acre parcel within our campus to our on-site power partner. Our on-site power partner would assume full responsibility for the entire power value chain on this footprint, including equipment procurement, environmental permitting, financing, construction, ownership, and operation of the generation assets.”

This confirms NUAI is not trying to partially own and operate the power stack. The on-site power partner sits on its own dedicated parcel, finances and operates its own generation assets, and controls the full power value chain on that footprint. That de-risks the project from NUAI’s standpoint and makes the real estate and campus development business easier to underwrite on its own (confirming how I had been modeling NUAI’s equity stake in the TCDC project).

Furthermore, management solidified their commitment to pursuing a NNN lease:

“We intend to commercialize our portfolio through long-term, triple-net lease agreements structured to deliver a targeted yield on cost. This pricing mechanism is designed to protect the Company’s operating margins by passing through all operating expenses, including taxes, insurance, and routine maintenance, directly to the tenant. Critically, our leases will treat power consumption as a pass-through expense or a direct tenant obligation, insulating our balance sheet from volatility in commodity fuel pricing.”

Here, management is confirming the model is long-term triple-net, yield-on-cost driven, with power largely pushed through to the tenant. That matters because it means NUAI is not trying to speculate on fuel spreads or absorb commodity volatility at the corporate level. The goal is to build the shell (in the real-estate silo) and monetize it through near 100% gross margin NNN lease revenues while keeping the operating exposure limited.

Put simply, management is telling you this should look more like infrastructure real estate with dedicated power attached, not like a merchant energy bet as some had been speculating.

Finally, I believe that the first signs of NUAI’s flywheel appear to be taking shape:

“We are in the process of leveraging industry-leading specialists across engineering and commissioning, on-site power generation and fiber and network capabilities. Together, with the site fundamentals of TCDC’s adjacency to natural gas, fiber, and CO2 infrastructure, we believe we are able to execute speed-to-power in a manner our peers can’t replicate.”

“We have cultivated strong, collaborative relationships with the Odessa municipality, positioning New Era as a preferred partner for regional infrastructure development.”

“We have relationships with leading private water infrastructure providers to source industrial and brackish water for our cooling systems. By utilizing non-potable water sources, we mitigate scarcity risks and align our operations with community sustainability goals, ensuring long-term viability and local stakeholder support.”

My take here is that this is NUAI’s competitive advantage and key differentiator.

Not a moat (yet) but the initial foundation of what could be one down the road.

As they build stronger relationships with these vendors, municipalities, and infrastructure partners, they should become easier to work with on future projects. That should improve execution, bring more opportunities their way, feed more deal flow back to those same partners, and further reinforce the value of partnering with NUAI versus a less connected developer.

In other words, the advantage is not just the land, but rather the developer ecosystem NUAI is forming around the land.

Liquidity and Risk

To be blunt, NUAI’s liquidity position is not pretty, at all.

Management says there is “substantial doubt” about NUAI’s ability to continue as a going concern. As of December 31, 2025, the company had just $1.2 million of cash, about $2.5 million of working capital, and expects to require roughly $73.9 million over the next twelve months, including $9.85 million by March 31, 2026 and up to $50 million by June 30, 2026 tied to financing obligations.

That is the headline risk shareholders need to understand.

Furthermore, I believe the usage of the shelf and thus the dilution risk to be real but should ideally be limited. This is what management had to say about their capital needs for the upcoming year and what the sources for said capital would be:

“We currently estimate that the total capital expenditures we will incur to complete the development of our flagship project could exceed $15 billion, excluding amounts expected to be financed by our tenants of which approximately $50 million to $300 million is expected to be incurred in the next twelve months across all phases. These near-term expenditures are expected to be funded through a combination of tenant prepayments, project-level debt financing, and strategic equity capital.”

Accordingly, I updated my model by increasing CapEx assumptions to $12.1B up from $9.5B for the real estate silo. I still exclude tenant fit-out and the power partner’s generation assets, but I no longer think my prior shell-only build cost was conservative enough for a true base case.

This adjustment brings my base case price target to $25.44.

The company is a venture-like public markets investment. Its core AI/HPC business has yet to generate revenue, and thus it still needs capital. It still has not yet signed binding tenant leases.

I believe that if NUAI gets the hyperscaler FID, that should unlock full access to both debt and equity capital markets for TCDC, with much of the equity funded by outside capital partners.

Accordingly, in that case, any additional dilution to NUAI shareholders should be limited outside of the SharonAI buyout and warrants, both of which I have already fully assumed in my model.

As investors, this should be understood, acknowledged, and accepted if you desire to possibly gain exposure to venture-like upside in the public market.

The market already knew this was an execution story. The company was never going to get from here to a 1+ GW campus without financing risk, tenant risk, and development risk.

All of these risks tie back to the big 4 hyperscaler FID. That is why this announcement, expected by the end of Q1, is the single most important event for NUAI in the firm’s history.

If the company is able to secure the FID, all other risk dominoes will fall into place. If the company is not able to secure the FID, shareholders have a very hard road ahead.

Final Takeaways

This was not a “good” 10-K in the traditional sense.

It was messy. It was painful. And it contained exactly the kinds of items you would expect from a company abandoning one strategy and aggressively funding another.

But I also do not think it was bearish, considering the core thesis revolves entirely around TCDC, not NUAI’s legacy assets.

Now, I will simply list some facts about next week’s business update and earnings call.

1. Under SEC rules, a company has to file an 8-K within four business days of a material event.

2. If management signed a major definitive agreement on Wednesday, March 11, say a hyperscaler lease, a power agreement, or some other material project contract, that four-business-day clock would expire precisely on Tuesday, March 17.

3. Tuesday, March 17 is the day NUAI has its 5:00 p.m. business update call.

Obviously, this does not prove anything. Tuesday’s call could simply be a routine follow-up to help the market digest the 10-K.

But the timing is very hard to ignore.

Furthermore, releasing your 10-K and then waiting a full four business days to complete your earnings call is highly unusual.

Typically, 10-Ks are released within a day or two of the earnings call.

My speculative take is that the 10-K was the final flush-out of NUAI’s legacy asset base and the full admission of all the “hair” around the company.

No more skeletons in the closet. The market now has all of NUAI’s less than commendable financials, liquidity situation, dilution overhang, and legal risk.

By giving the market four full trading days to digest that news and price it in, management is giving investors a fresh palette for whatever Tuesday’s call holds.

Godspeed.

View my full NUAI model and price target here.

Disclosure: I am long NUAI. I may buy, sell, or change my position at any time without notice. This article reflects my personal opinions and is for informational and educational purposes only. Nothing in this article should be construed as investment advice or a recommendation to buy or sell any security. I have not received compensation from NUAI, its management, or any affiliated party for this content. I have no business relationship with the company.

Intro

NUAI’s 10-K was announced yesterday (3/11) and should hit EDGAR this morning. With the earnings call and business update set for March 17 at 5pm ET, the pieces appear to be falling into place for a major FID announcement. At the very least, investors should get meaningful color on project progress as management approaches its Q1 FID deadline.

While investors eagerly wait for the official announcement, the writing is already on the wall. More specifically, the writing is on the Odessa/Midland job boards.

If you remember, management said on the BofA call that they would be announcing a new development partner, presumably the group that would actually handle the buildout of the powered shell.

First, it is worth level-setting. There are only two meaningful data center-related projects in Odessa, Texas, where TCDC sits.

One is TCDC: a 1 GW hyperscale data center campus designed around advanced liquid cooling, modular construction, and behind-the-meter power generation, all purpose-built for the latest AI chips.

The only other real data center project in the area is Cipher Mining’s site. But that site is already built. Cipher broke ground in 2022 and has been actively mining Bitcoin there for years.

More importantly, on Cipher’s recent earnings call, analysts explicitly asked CEO Tyler Page whether the company planned to convert its Odessa site from Bitcoin mining into an HPC/AI data center. Page shut that down, stating there are “no immediate plans to convert the site to HPC.”

Cipher is not building a new hyperscale AI campus in Ector County. So if a contractor is staffing up for a landmark hyperscale campus in Odessa, there is really only one logical candidate: NUAI’s TCDC project.

Zooming out a bit, Midland matters here too. For those unfamiliar with the area, Midland is where much of the white-collar oil and gas infrastructure sits, while Odessa is where much of the blue-collar field activity, work yards, and temporary housing are based. The two cities are only about 20 miles apart, share an airport, and effectively function as one integrated regional hub for West Texas industry.

Full credit goes to Stocktwits user TheSterl for surfacing the initial GCON connection.

And if you look around the broader Midland/Odessa region for other projects that could plausibly explain specialized AI/data center hiring, the list is thin. There are plenty of blue-collar training yards and traditional industrial projects, but nothing remotely comparable to the scale or technical demands of TCDC’s 1 GW campus.

Accordingly, the logical next step is to see who is hiring, where they are hiring, and what exactly those roles are describing. Once you do that, a pretty interesting picture starts to emerge.

Over the last month, a perfect storm of firms has posted highly relevant job listings. Broadly speaking, they fall into three buckets:

Those building the shell: GCON (Odessa)

A specialized general contractor backed by Japanese giant Obayashi that focuses on physical construction and powered-shell infrastructure.

Those building the power generation: Gemma (Odessa)

A premier industrial EPC firm focused on large-scale thermal power generation and complex power-train buildouts.

Those handling the money: Turner & Townsend (Midland)

A CBRE-owned global consultancy that provides institutional-grade cost management, quantity surveying, and commercial governance for massive infrastructure programs.

Put simply, you have one cluster building the campus, one cluster building the power, and one cluster putting financial and commercial controls around the entire thing.

That is the exact ecosystem you would expect to see if TCDC were moving from blueprint to reality.

Building the Shell: GCON

The clearest green flag here is GCON.

GCON is currently hiring for 13 highly specific roles in Odessa, and the job descriptions scream TCDC. In my view, they are simply incompatible with any other project currently underway in the region.

Across all 13 postings, GCON includes the following header:

“GCON is delivering a landmark hyperscale data center campus in West Texas, purpose-built to support next-generation AI, cloud computing, and high-performance workloads.”

Again, there is really only one project in Ector County that fits that description.

And when you dig into the individual roles, it gets even more specific.

Underground: Superintendent – Underground Utilities, Field Engineer – Civil

These roles are tasked with executing “duct banks, manholes, storm, sanitary, and site utilities critical to AI facility operations.” That language matters.

Duct banks are essentially the concrete highways for high-voltage power. You don’t pour those for a warehouse, basic industrial facility, or some generic local server room.

One of the postings even describes this scope as the work that “establishes the foundation for all vertical and MEP construction that follows.”

In other words, this is the enabling infrastructure for a major campus.

Power & Liquid Cooling: MEP Superintendent, MEP Project Manager, Electrical Superintendent

These roles mention facilities with “extreme power density and advanced cooling requirements” and oversight of “MV/LV power, UPS, generators, and distribution systems.”

That is not standard data center language.

The references to extreme density and advanced cooling strongly suggest a design tailored to AI workloads, where rack densities and cooling requirements are dramatically higher than in traditional enterprise deployments. And the inclusion of UPS systems is important too. That implies an enterprise-grade, mission-critical environment where even milliseconds of downtime matter.

That is hyperscaler territory.

Sr. Project Manager, Project Manager, Project Engineer

These roles are described as part of a “2–5 year, project-based assignment” spanning “multiple phases and buildings—not a short-duration project.”

That lines up very cleanly with NUAI’s plan for a multi-phase build.

Commissioning Manager, Quality Control Manager, Safety Manager, Safety Coordinator

The Commissioning Manager role may be the most telling of the bunch. It references leading “testing, validation, and turnover... to meet hyperscale AI performance standards.”

There is not much ambiguity there.

And the fact that GCON is hiring both early-stage construction roles and late-stage commissioning roles at the same time suggests a highly compressed delivery schedule. That would make perfect sense if the project is being modularized, which management has already indicated is part of the strategy.

Project Accountant, Project Coordinator

Even the finance-side roles matter. These postings reference “project-specific billing, subcontractor buyout, and change management.”

Put it all together, and GCON looks very likely to be the development partner NUAI alluded to on the BofA call.

Building the Power Generation: Gemma

If GCON is the tell on the data center shell, Gemma is the tell on the power plant.

And if you want proof that the 450 MW natural gas buildout NUAI has been talking about is not just conceptual, Gemma’s Odessa postings are about as close as you can get without an official press release.

OEM Manager

This role is especially revealing.

The posting describes the OEM Manager as the primary interface between Gemma and the “Major Equipment / Power Train Supplier” and specifically references oversight of “all OEM heavy-haul deliveries – including rail siding.”

On February 27, NUAI announced that it had secured 450 MW of power generation through Turbine-X. You don’t need heavy-haul rail logistics for data center servers. You need that for industrial-scale generation equipment like gas turbines and associated power-train hardware.

The same posting also references ensuring “Ready for First Fire status of each powertrain.”

That is not vague construction language. “First Fire” is the term used for the initial ignition of a gas plant.

In other words, Gemma is hiring the person whose job is to manage delivery and commissioning coordination for the BTM power-generation backbone of the site.

Quality Inspector

This posting requires experience with “simple-cycle, combined-cycle, or other large-scale thermal power projects” and calls for an “Immediate” start.

The role also references “soil testing, concrete testing, and hydrostatic testing,” which points to active foundational work on a thermal generation asset.

Again, not theoretical. Not future-state. Immediate.

Assistant Project Manager (APM)

This role is responsible for “negotiating & executing project contracts” and participating in “monthly updates to the project cost forecasts,” while also serving as the “onsite representative... effectively communicating with Owner.”

That sounds like a live project with an active owner relationship, active forecasting, and active contract execution.

Now, to be fair, Gemma is also the lead contractor on the CPV Basin Ranch project in nearby Ward County. So that is the obvious alternative explanation.

But I don’t think that explanation holds up particularly well for three reasons.

1. Timing: CPV Basin Ranch is a longer-dated grid project with completion targeted around 2028–2029. By contrast, these Gemma roles are for immediate mobilization in Odessa.

2. Vocabulary: This is a subtle tell, but I think it matters. In large grid-scale power projects, equipment is more commonly described as a “Power Block” or “Turbine Island.” Here, Gemma is using the term “Power Train,” which aligns much more naturally with modular, integrated engine-and-generator configurations like those associated with Turbine-X.

3. Geography: These postings are centered on Odessa, which is where TCDC sits.

Taken together, I think the odds strongly favor TCDC being the client behind these roles.

Handling the Money: Turner & Townsend

If GCON signals the shell and Gemma signals the power, Turner & Townsend signals something just as important: institutionalization.

T&T is a global consultancy majority-owned by CBRE, and CBRE manages the global data center footprints of the Big Four hyperscalers: Microsoft, Google, Meta, and Amazon.

That does not prove anything on its own, of course.

But it definitely matters that T&T is hiring in Midland for a role called Senior Cost Manager / Quantity Surveyor – Data Center Construction.

The job description references providing “commercial rigor and governance” and managing the “Bill of Quantities” for a “large-scale, mission-critical data center program.”

A Quantity Surveyor is not some generic administrative role. This is the type of professional brought in to measure, audit, and financially control every major cost line item on a very large project, from copper and concrete to subcontractor claims and change orders.

You only hire that person when the project is large enough, expensive enough, and institutionally important enough that cost control itself becomes a core workstream.

The phrase “mission-critical” matters too. In infrastructure and construction circles, that is essentially code for environments where uptime is non-negotiable, namely data centers and power systems that cannot fail.

Put simply, Turner & Townsend does not get brought in because someone has an idea.

They get brought in when the idea is turning into a multi-billion-dollar program that needs real governance.

In my opinion, their presence in Midland may be one of the strongest signals we have that this is no longer just a startup vision. It suggests the project is being structured, monitored, and costed in a way that aligns with real institutional capital and real anchor-tenant requirements.

That does not necessarily mean FID is signed, but it does suggest the project is much further along than the market is giving it credit for.

Supply Chain and Systems Integration

The three buckets above are the core tell. But the supporting cast matters too.

A handful of other postings from global EPCs and specialized service firms add another layer of confirmation that the power systems, controls, and commissioning stack is actively being assembled.

Hanwha Ocean: Construction & Commissioning Manager - Power Plant

Hanwha has postings in the Midland/Odessa area for both a Construction Manager and a Commissioning Manager tied to “gas turbine power plants... and data centre power plants.”

The listings also reference ensuring “Ready for First Fire status of each powertrain,” which again ties directly into gas-turbine commissioning language.

These are the people you hire when the hardware is real, when the systems have to be tested, and when someone eventually has to flip the switch.

This may be the single clearest third-party listing in West Texas that explicitly links gas turbines with a data center power plant.

And that is essentially the TCDC blueprint in one sentence.

Saulsbury Industries: Assistant Project Manager - Plant & Buyer

Saulsbury is a local Odessa staple on major industrial projects. Seeing them hire an APM for a plant and a buyer for field services suggests active procurement and active field execution.

In plain English, someone is ordering pipe, valves, steel, and all the other unglamorous but necessary inputs that precede major site work.

System One: Controls Specialist - Data Center

This one is particularly interesting because it points toward the power-and-data-center handshake.

The role references PLC-based control systems for “Chillers, Switchgear, and Backup Generators.” PLCs are the industrial computers that make all of these systems talk to each other.

That matters because the campus is not just a data center and not just a power plant. It is an integrated system. Someone has to make the cooling, switching, generation, and redundancy architecture all work together in a stable and automated way.

This is exactly the kind of role you would expect to see once the project moves from concept drawings to systems integration.

Bird Electric (subsidiary of Black & Veatch): Project Accounting Manager

Bird Electric is hiring for project accounting with a focus on “percent complete accounting.”

In major construction, percent-complete accounting is how contractors recognize revenue as milestones are achieved. You don’t use that framework when nothing is happening. You use it when a project is physically moving and contractors are billing against real progress.

So if someone in Midland is hiring around this function right now, it strongly suggests there is an active contract structure and active milestone-based execution underway.

Final Thoughts

None of this is the official FID announcement.

But that is exactly the point.

The market keeps waiting for a clean, formal press release, while the real-world evidence keeps piling up in plain sight.

You have a specialized hyperscale contractor hiring in Odessa for a “landmark hyperscale data center campus in West Texas.” You have a power EPC hiring around “power trains,” “heavy-haul deliveries,” and “Ready for First Fire” status. You have institutional cost-control professionals being brought in to govern a “large-scale, mission-critical data center program.” And then you have supporting roles across commissioning, controls, procurement, and project accounting that all fit the same technical blueprint.

Individually, maybe each datapoint could be explained away.

Collectively, that starts to get difficult.

Put differently, this is what a project looks like when it is moving from blueprint to execution.

Put simply, the writing is already on the wall.

And in this case, the writing is on the Odessa and Midland job boards.

Disclosure: I am long NUAI. I may buy, sell, or change my position at any time without notice. This article reflects my personal opinions and is for informational and educational purposes only. Nothing in this article should be construed as investment advice or a recommendation to buy or sell any security. I have not received compensation from NUAI, its management, or any affiliated party for this content. I have no business relationship with the company.

This is an update to my last NUAI deep dive (posted March 5, 2026), which you can read here.

Since my last post on NUAI, Why NUAI Has 10x Potential From Here, I have compiled bits and pieces of research, thoughts, and takes on the company.

This will be a short and sweet postscript to my last deep dive that also covers my insights from CEO Will Gray’s interview at PTC ‘26, which you can view here.

What is the downside case for NUAI?

I have spent a decent amount of time thinking about this in recent days. We know that the reward is there if NUAI lands the Big 4 hyperscaler FID, but what is the risk for shareholders if, say, the hyperscaler backs out for some unforeseen reason?

As Charlie Munger loved to say:

“Invert, always invert”

So, what I will ask is: how could this deal get blown up, and what would NUAI be worth in that case?

1. The FID simply gets pushed back / delayed.

   1. We are expecting a FID signature by March 31. With a $10M payment due to Sharon AI that same day and a $50M note maturing in June, the hyperscaler could slow-walk negotiations to squeeze a micro-cap for better terms.

2. The ongoing soil sampling comes back negative and/or results in months of delays due to some contamination or waste being present.

   1. The February 26 PR confirmed ongoing sampling and “abandoned pipeline” removal. If the study reveals a major environmental plume or high-pressure orphan lines not on the maps, the 2027 power timeline will be delayed.

3. Hyperscaler CapEx declines rapidly, and thus they get cold feet and back out of the deal.

   1. If GPU demand cools or the market punishes massive infrastructure spending, the GW-scale project may be put under indefinite review.

4. Capital dries up due to macro conditions (Iran).

   1. This thesis relies on leverage via large amounts of debt financing. Geopolitical shocks in the Middle East or 10-year yields spiking would slow the capital markets required to make the $10B campus pencil.

Though I feel these scenarios are unlikely, they are 100% possible. Accordingly, we should consider what NUAI would be worth in such a bear case.

In the BofA call, President & COO Charlie Nelson specifically said that the purchase price for SharonAI’s 50% stake was based on the value of the powered land.

“But through that time the land became powered land by nature of all the work that we had put into it and in co-developing it. And so, over the course of this time we’ve had multiple bidders come in and try to buy the land from us, so we had kind of a relative peg on what the land is worth. So it wasn’t quite to powered land pricing, so we just essentially agreed on a structure – and it wasn’t just like a $70 million cash slog up front, it comes in waves based on milestones, etc., all the structure is publicly announced, so that’s pretty easy to find – but it was really determinant upon what the land is worth as powered land was the intention, because that’s a fair way to price it if you’re looking to buy them out.”

Based on the $70M Sharon AI buyout structure, we can assume the total value of the land is $140M ($70M x 2). This would equate to a $/MW metric of $560,000/MW.

If management was willing to pay $70M for 50% of the project based on powered land pricing (backed by unsolicited bids, might I add) back when NUAI only had 250MW under LOI, we have a clear peg for what NUAI’s land could be worth today.

Since then, NUAI has secured 450MW of BTM turbines from Thunderhead. If we assume the same $/MW metric, we net out to a valuation of $252M.

NUAI has total outstanding liabilities of $60.49 million ($50M note to SharonAI, $2.36M ARO, $2.16M taxes, and $5.9M in other liabilities).

Netting the liabilities from the asset value ($252M - $60.49M) gives us a valuation of $191.51M.

Using a highly conservative fully diluted share count of 88.2M, this would equate to a per-share valuation of $2.17.

Given NUAI’s current share price of $4.59, that gives us downside risk of 52.72%.

Given my updated model’s base-case target of $27.26, that gives us upside of 494%.

Thus, our risk-reward on NUAI is 9.37:1 (494:52.72).

Obviously, there are many factors at play here, but it is an interesting thought exercise.

Who is the hyperscaler?

This is a speculative theory, but there is mounting evidence that the hyperscaler NUAI is currently negotiating with is none other than Microsoft.

For months, the TCDC site lived under the local radar as Project Thunderstorm (a nod to NUAI’s power partner, Thunderhead). But on September 11, 2025, during an Odessa Development Corporation (ODC) meeting, the project’s name was changed to Project Blizzard King. As the Texas Governor’s Office stepped in to lead the incentive negotiations, the projected CapEx was officially revised from a local $100M+ build to a staggering $30B+ mega-campus.

State-level codenames are rarely random; they are almost always a subtle tell related to the parent company, often utilizing a subsidiary’s name to maintain a layer of privacy.

Microsoft completed its $75.4 billion acquisition of Activision Blizzard (which included the mobile giant King) on October 13, 2023. In Microsoft’s official financial and SEC filings, the company now explicitly organizes the unit as Activision Blizzard King (ABK).

The connection goes deeper than just a name. Primary Digital Infrastructure (PDI), NUAI’s partner for TCDC, is a key member of the JV developing the Stargate Project in Abilene, TX, as in the Stargate project that is backed by Microsoft and OpenAI.

You can decide what to take away from these facts on your own.

What about ESG?

So how is NUAI supposed to convince a Big 4 hyperscaler to abandon its ESG commitments (which, to my surprise, they are still taking very seriously despite the current administration’s lack of regard for ESG)?

Put simply, they don’t. NUAI had planned for this hurdle from the first month or two of their pivot. Let’s take a look at two members of their board: P.J. Lee (co-founder of Enfinity Global, a 20+ GW renewable developer) and Trent Yang (Managing Director at Galway Sustainable Capital).

Now why in the world would NUAI have added two highly experienced green finance individuals to its board? My theory is that they have been planning to offer a couple of things to their end tenant, but the main options are 1. carbon capture and 2. virtual power purchase agreements (VPPA). While we very well may see carbon capture done at TCDC, I also believe that we will see a VPPA.

Now what is a VPPA, why would a hyperscaler want one, and why am I so confident that NUAI will offer one to its end tenant?

How it works: Hyperscalers match their on-site natural gas consumption (24/7 reliability) by purchasing Renewable Energy Certificates (RECs) from off-site solar/wind farms. Coincidentally, Enfinity Global is a massive U.S. solar developer.

Precedent: Meta’s Hyperion Project (Louisiana) and Microsoft’s Texas ERCOT facilities utilize this exact REC-matching playbook to claim to be 100% renewable while physically running on gas.

Another offering that NUAI could work with the end tenant to develop (again, if desired — Will Gray has been very clear that NUAI is building to the hyperscaler’s spec) is an on-site wastewater treatment facility.

In the Permian, for every 1 barrel of oil produced, approximately 10 barrels of produced water (industrial wastewater) are generated. In Will Gray’s recent interview at PTC, he mentioned that the processing of this wastewater for use in data centers was something they were looking into:

“So for instance, oil wells produce a lot of water... if you produce a barrel of oil, you may get 10 barrels of water. So what they're doing now is they're actually recycling the water, repurposing it, so you can actually use this for data center purposes. It's just recycle... So as opposed to using municipality water, they’re using [this]. In fact, I think it's quite imperative understanding... utilizing wasteful gas and also water on top of that.”

This would eliminate the need for municipal/drinking water, turning local waste into a valuable asset.

Last but not least, tons of gas gets flared off in the Permian, releasing CO2 into the atmosphere without any utility. NUAI is seeking to use as much of this flare gas as possible, again generating value out of what would otherwise have been waste.

What about the data center labor shortage issue?

For reference, the Abilene Stargate site currently employs over 6,400 construction workers. Around the country, we are seeing data center projects struggle to field the absolutely massive number of skilled blue-collar laborers needed to put these data centers together.

According to the Associated Builders and Contractors (ABC), the U.S. construction industry faces a shortage of roughly 439,000 workers in 2026. This is expected to climb to 456,000 in 2027.

Anecdotally, many executives are calling the labor shortage in data center buildouts right now one of the most pressing bottlenecks to construction.

When asked how NUAI was planning to tackle this issue, Will Gray stated that they are targeting 80-85% modularization (building the data center off-site in chunks) to bypass the local labor shortage and reduce on-site talent needs to just 15-20%.

Furthermore, local communities such as Abilene, which host these massive data center buildouts, experience massive spikes in rent as all of these workers flood the town. New housing and infrastructure get built, but once the project is complete, it leaves a sort of ghost town with far too much infrastructure for the permanent residents.

I am a bit skeptical of the efficacy of such a modularity strategy (80-85% feels really high), but if possible, this would solve yet another pressing bottleneck facing data center construction.

Coincidentally, Microsoft is a leader in this modular buildout methodology. Their new Fairwater AI data center design (used in their Atlanta and Wisconsin superfactories) is built around modular, repeatable blocks. They use prefabricated power and cooling skids that are built in a factory, tested, and then plugged in on-site.

Reports from early 2026 show that 40% of data center operators have increased their investment in prefabricated modules.

What will NUAI look like 10 years from now if successful?

When asked this exact question in his PTC interview, Will Gray stated:

“I think it's like any large industry, there will be a consolidation. So, it's either you get consolidated or you are the consolidator. And we want to be the consolidator because again with our energy backgrounds, we think it's a natural fit.”

"Do we grow through organic growth or through M&A? And I think it's going to be attributable to both."

NUAI isn’t trying to build one data center, but rather a blueprint that it can take onto its nth project. The idea of NUAI growing through M&A is appealing to me as someone working in PE, but before that, they obviously have to land this FID for project #1.

The growth potential is there, the downside is decently protected, we nearly have 10:1 risk-reward asymmetry, and there is a real reason to believe Microsoft could be who NUAI is negotiating with.

No company is perfect, and without a doubt there is massive risk here, but again, I know of no other company as asymmetric as NUAI at a $250M market cap.

View my full NUAI base-case DCF model here: Link

Disclosure: I am long NUAI. I may buy, sell, or change my position at any time without notice. This article reflects my personal opinions and is for informational and educational purposes only. Nothing in this article should be construed as investment advice or a recommendation to buy or sell any security. I have not received compensation from NUAI, its management, or any affiliated party for this content. I have no business relationship with the company.

This is an update to my original NUAI deep dive (posted Jan 27, 2026) that you can read here.

Following the Bank of America conference call with NUAI’s executive team last Friday (the recording just came out today and you can listen here), the fundamental mechanics of the TCDC 1 GW campus infrastructure have changed entirely.

I was wrong about my previous thesis. They aren’t going to acquire the neighboring Quail Run power plant (my apologies for the false ideation). President and COO Charlie Nelson was blunt about why that legacy steel doesn’t fit the mission, noting that “combined cycles in particular are not great to complement with data centers, because combined cycles, you can’t turn down as easily.”

Instead of a sluggish industrial furnace, they’re going with simple-cycle aeroderivatives (jet engines) that “can throttle it down, you can throttle it up, and that happens very, very quickly.” What they are actually doing is significantly leaner, more capital-efficient, and mathematically much more lucrative for shareholders.

The physics behind why NUAI is choosing jet engines over traditional plants is intriguing and worth understanding. To do so you have to understand the “jagged load” of AI. Unlike a Netflix server that streams data smoothly, an AI training cluster of 50,000 GPUs works in synchronous pulses.

Think of it as a massive rowing crew. All 50,000 GPUs pull at once to calculate math (a massive 1 GW spike), and then they all stop at once to “talk” to each other and synchronize what they learned. In those few milliseconds of “talking,” the power draw craters.

This creates a “sawtooth” power profile that would stall a traditional combined-cycle plant. If you slam the brakes on a sluggish industrial furnace, it can’t recover fast enough for the next spike. NUAI’s simple-cycle aeroderivatives are the only thing that can throttle fast enough to mirror the jagged heartbeat of an AI model.

NUAI is pivoting from a capital-heavy power owner to an asset-lite, pure-play BTM powered data center developer platform. By offloading the heavy machinery and commodity risk to their partners, they’ve engineered a way to deliver a $10-15B custom hyperscaler-ready Tier-1 data center shell while radically slashing their own equity requirements.

Here is the full, bottom-up deep dive into exactly how these changes affect the DCF model, the data backing every assumption, and why my base case (75% probability of success) price target now sits at $30.51.

1. The Decoupling of the Shell and Power Infrastructure

Previously, I had assumed there would be a single entity collecting a massive, all-in check from the hyperscaler. Now, we see a bifurcated power and shell structure.

The Power Silo (Thunderhead & Harbert)

The power generation assets (those simple-cycle jet engines) sit in a separate silo that NUAI does not have a stake in. Harbert Management Corporation (HMC) and Thunderhead are footing the bill for the engines and they will generate their own IRR (return on investment) through a direct contract with the hyperscaler. Charlie Nelson broke this PPA (Power Purchase Agreement) into three distinct buckets paid directly to the power partners:

1. Reservation Charge: A capital recovery fee that pays for the turbines.

2. Operating Fee: Covers the actual maintenance and engine mechanics.

3. Gas Component: A pass-through of the fuel cost.

By staying out of this silo, NUAI avoids the massive CapEx, financial exposure to fuel price volatility, and the mechanical risk of running a 1 GW power plant. Harbert and Thunderhead get paid from the electrons; NUAI gets paid from the real estate.

The Real Estate Silo (NUAI and Partners)

The call confirmed that NUAI’s monetization is now clean NNN (triple net lease) rent. In layman’s terms this basically means that all costs (utilities, insurance, and all other OpEx) are passed directly onto the tenants. NUAI owns the site, the powered land status, and the 1 GW of liquid-cooled shells. The hyperscaler pays NUAI a lease purely for the shell, the cooling infrastructure, and the physical security of the site.

This is a much higher-quality revenue stream. It is static, predictable, and carries near-100% margins because the variable costs (fuel and turbine maintenance) are siloed away with Harbert and Thunderhead. This structure allows NUAI to act as the Master Developer collecting a high-margin toll on the most valuable piece of dirt in the Permian while letting the power specialists carry the heavy metal.

TL;DR: NUAI moves from a revenue stream that was exposed to the volatility of gas prices and extraneous utility costs to a highly predictable, contracted, near 100% margin cash flow. Wall Street loves this shift because they can place high levels of confidence in future cash flows, which reduces discount rates and brings the present value of the company dramatically higher when NUAI’s first revenues are 2+ years out.

2. The Model (Real Estate Silo Only)

Here is how I am thinking about modeling NUAI’s DCF from a birds-eye view. I am primarily concerned with TCDC (obviously) and am giving NUAI no credit for its New Mexico land, given the ongoing legal complexities in that state, nor any future deals if they are to progress to a full platform. I would argue that this bakes in a level of conservatism.

I use a real estate project based finance approach where you model a specific deal in totality then distill your claim on the project’s total value based on the % of the equity in the deal that you own. This is the #1 outstanding question for NUAI (how much equity they will get in the final project). As you will see later the model is highly sensitive to this input and accordingly shareholders should be laser focused on this figure if/when the FID comes out. Working with similar deals in the sector and assuming NUAI is able to leverage what they bring to the table (powered land and a hyperscaler anchor tenant) I feel comfortable modeling a 20% stake in the project’s equity. I could see this figure going down to 10% in the worse case scenario or as high as 30% in the best case scenario.

You might be wondering why NUAI wouldn’t walk away with more of the pie if they put this whole deal together and the answer is cash. As many followers have brought up time and time again NUAI as a micro cap simply does not have the cash to put down to fund a $10B+ project. Because of this NUAI has and will partner with capital partners and data center developers to actually build out the shell and bring the wheelbarrows of cash needed to complete the project. NUAI on the other hand gets sweat equity (GP interest/promote) for orchestrating the whole thing while having to put very little actual cash into the actual build.

One of the first steps is to figure out how much of the project will be funded by debt and how much will be funded by equity (cash). In this case I am assuming an industry standard for hyperscaler anchor tenants at 65% LTC (the % of debt used relative to the total project cost).

From there you project out how much CapEx (investment) you will need and exactly which years that capital will be called. In my model I followed management’s guidance from the call for when power will come online.

From here we can roughly estimate when the capital will be called (used).

As each use of capital will be partly funded with debt we can now figure out how much debt we will be taking out and when (giving us a debt schedule). For simplicity’s sake I am modeling a Payment in Kind (PIK) debt facility which basically means that interest is not paid upfront but rather just gets added to your total debt balance and is paid off once the project is sold. I also assumed that all proceeds from the data center would be swept (used) to pay down their debt balance. These are both mechanisms typically see in large scale debt funded projects like this in real estate or LBOs.

In terms of the revenue side of things I simply modeled a shell gross base rent figure based on industry comps for fully custom, water cooled, Tier 1 GW scale AI data centers.

Due to this being a real estate project I used a NOI / Cap rate valuation to determine what the full scale data center will be worth in 2030. Using our revenue and online power assumptions and near 100% margins we can back our way into a NOI of ~$1.98B. In this model I assume a terminal cap rate of 6% given the massive demand for data centers and the inherent scarcity the industry is experiencing. This nets us to a total valuation at sale or Total Enterprise Value (TEV) of ~$33B.

Due to the dynamics of the cash sweep paying down the debt over the life of the project and the value creation from taking a GW scale data center online the equity value massively outstrips the value of the debt at the sale of the asset. Assuming you get that $33B in cash you then can pay off your remaining debt (~$2.5B) and whatever you are left with ~$30.5B you can now distribute to the owners of the project.

This is where that 20% NUAI equity split assumption comes in. Because I model NUAI having a 20% claim on the project level equity NAUI walks away with 20% of that final equity value of $30.5B. This nets out to roughly $6B for NUAI in 2030. This sounds great but we have to discount this to the present day due to the time value of money (money now is worth more than money later).

Assuming a 10% discount rate and a 5 year hold time NUAI’s Net Present Value (NPV) in the project based on the model (what this project is worth to NUAI today) is ~$3.8B.

The final piece to this puzzle is figuring out how many shares to distribute this value between. My model assumes that the CEO and COO both get their full PSU stock grants, all warrants are executed, the SharonAI dilution, and an incremental 12M of shares from the shelf being sold. This gives us a final fully diluted 2030 share count of ~93M. Dividing that final NPV equity value between all of those shares and multiplying by the probability of success (75%) nets us a current price target of $30.51.

To get an idea of exactly how lucrative these data center projects can be for all participants given the massive demand and limited supply I also built a cash flow table (which tells you when cash leaves your hands and when it comes back) to calculate the Internal Rate of Return (IRR) or in other words how much this investment would return per year. Needless to say the returns can be astronomical.

Finally I built a sensitivity analysis to see what changes in the most critical assumptions would yield in terms of price targets for NUAI. Based on this, one could argue that the market is currently giving NUAI and TCDC a sub 20% implied odds of success. You can decide for yourself if the market is massively mis-rating NUAI.

You can view the full model here. If you are an NUAI investor, I highly suggest making a copy and playing around with the model’s assumptions.

3. The Unannounced Partner

I previously assumed Primary Digital Infrastructure (PDI) was providing the balance sheet given their partnership but Nelson clarified that PDI, despite having founders from CyrusOne and Digital Realty, is acting strictly as an advisor (aka a super fancy well connected Realtor). During the call Will Gray stated “At the request of the hyperscalers themselves... there’s a specific execution partner that wanted us to bring in... This is unannounced.”

A hyperscaler isn’t going to let a micro-cap manage the concrete pours for a $10B liquid-cooled shell. A Tier-1 mega-developer (think CyrusOne, QTS, or Vantage) is more than likely stepping in to provide the Engineering, Procurement, and Construction (EPC) muscle.

Side note: To hit a 65 percent LTC, or really any meaningful amount of debt on a 9.5 billion dollar build, you need a signature from one of the big four hyperscalers: Meta, Microsoft, Amazon, or Google. Whoever it is, the lender is looking at their credit worthiness, not NUAI’s. We saw this reality play out in the private credit markets recently when Blue Owl had trouble shopping debt for a CoreWeave facility. It served as a reminder that even in a red-hot AI market, credit quality still matters. By landing a big four anchor tenant, NUAI is essentially using those trillion-dollar balance sheets to back the project. It is the only way to get a deal of this magnitude across the finish line without having to use much lower leverage ratios (and thus more equity).

4. The Holy Trinity of Regulatory Engineering

During the call, Nelson mentioned that the first phase of the build-out (200 MW) will be “grid-ish” using a bidirectional interconnect by late 2027. By registering as a generator capable of pushing power back to the grid, NUAI bypasses the queue.

Now you, like me when I first heard this, might be wondering how in the world NUAI is planning to get connected to the grid in 2027 when ERCOT queues are 5-7 years long. The answer is honestly a stroke of genius from NUAI management. They are taking advantage of 3 different regulatory loopholes to jump the line entirely.

First, they are standing in a different line. Because ERCOT is in desperate need of dispatchable power, the process for registering as a generation resource is fundamentally different than applying as a massive data center load. Because NUAI is going to install 450 MW of engines, they legally qualify as a generator.

Second, they are using existing physical infrastructure to shorten the paperwork. ERCOT allows for “Limited Operations” or “Interim Interconnection” for generators that are physically adjacent to massive switchyards with existing capacity. This is exactly why locking down the 54-acre strip that connected NUAI to the two legacy power plants was so critical to the hyperscaler in negotiations. Because the transmission lines are already built, NUAI avoids the standard five-year queue for new transmission planning. Instead, they trigger a localized study process which can be completed in 12 to 18 months.

Finally, they are offering the grid exactly what it needs to ensure those applications get approved. As Nelson pointed out on the call:

"By registering as a dispatchable resource under the new Texas energy framework, we aren't just a tenant; we are a solution to the grid's frequency stability issues."

Because NUAI’s jet engines can throttle up and down very quickly, they qualify as Fast Frequency Response (FFR) assets that can actively help prevent blackouts.

By having a bidirectional interconnect, the site can physically push power back onto the grid during high-demand events like a Texas freeze, and pull a small amount of startup power or maintenance power in the meantime. This startup power is exactly what NUAI plans to use to run the 200 MW Phase 1.

NUAI is effectively cloaking their 200 MW Phase 1 data center load behind a generation registration. It is an absolute masterclass in regulatory engineering.

5. De-Risking the Permian

The final pillars of Permian Basin risk are water scarcity and environmental optics. If you want a big four hyperscaler to sign a 20-year lease, you have to solve both. The BofA call confirmed that NUAI has thoughtfully neutralized these liabilities.

Water

Traditional data centers consume millions of gallons of water for cooling, which is an absolute non-starter in the West Texas desert. Nelson confirmed they are bypassing this entirely by utilizing air-assisted liquid cooling in a closed loop. Instead of evaporating millions of gallons of local groundwater, this closed-loop system brings consumption down to under 10,000 gallons.

ESG

Hyperscalers are strictly bound by corporate ESG targets. Usually, building a massive gas-powered data center would trigger an environmental PR nightmare. However, Will Gray highlighted that this project is actually solving a local ecological disaster. By capturing stranded gas that is just flared off, NUAI is converting waste emissions into high-density compute.

Gas Security & Expansion

The ultimate limitation for any off-grid power project is fuel delivery. On the call, Will Gray pointed out a fatal flaw in most BTM data center pitches. Many developers secure land with access to only a single gas pipeline. That represents a massive single point of failure because pipelines inevitably face routine maintenance, pressure curtailments, or mandatory anomaly repairs. If a defect in the single pipeline is found and your only fuel source is shut down for a three-day repair, your gigawatt campus goes dark. For an AI cluster requiring 99.999% uptime, a three-day outage is a catastrophic and unacceptable failure.

The Ector County site solves this by sitting directly on an intersection of three major midstream pipelines: ONEOK, Enterprise, and Whistler. This physical convergence provides the absolute redundancy a hyperscaler requires for guaranteed uptime. Beyond just keeping the lights on, it means the infrastructure isn’t capped at the initial 1 GW. According to the call it has the underlying fuel logistics to support north of 2 GW, giving the anchor tenant a massive runway for future expansion without ever needing to permit a new site.

6. Macro Grid Tailwinds

This off-grid, behind-the-meter (BTM) structure is rapidly becoming the federally preferred blueprint. As Nelson pointed out regarding the shifting landscape, “Behind the meter is not the exemption, it’s now becoming the rule.”

But beyond the regulatory tailwinds, the true catalyst for BTM adoption is a massive, structural convergence in the Total Cost of Ownership (TCO). Historically, building your own power plant was seen as a hyper-expensive bridge until cheaper grid power arrived. The BofA call shattered that assumption, revealing that the cost of BTM power and grid power are actively converging, and in many cases, grid power is becoming the more expensive option.

To understand this convergence, you have to look at how grid power is priced. Three or four years ago, a massive industrial user in Texas could buy power off the grid for roughly $45 per MWh. Today, those open bids are sitting between $85 and $95 per MWh.

Why the massive spike? The answer is Transmission and Distribution (T&D) costs. ERCOT is currently buckling under a 230 GW backlog of demand. To handle even a fraction of that, the state is forced to build massive new 765 kV transmission systems and execute broad grid upgrades. The cost of those multi-billion-dollar infrastructure projects is shared collectively by everyone pulling off the grid via escalating T&D charges. As Nelson bluntly stated, “The only answer to stress on the grid is expansion of the grid. Expansion of the grid means higher T&D charges, period.”

This is where NUAI’s model shines. Because NUAI’s power is generated locally BTM, the tenant bypasses these shared T&D taxes entirely.

While the actual generation hardware might be slightly more expensive upfront, the absence of T&D fees brings the localized power cost to near-parity with the grid. More importantly, it creates a predictable, static cost structure. A hyperscaler can lock in their base equipment charge and hedge their Waha gas supply for years, creating a flat operating expense. Meanwhile, a competitor relying on the grid is exposed to runaway, unpredictable T&D inflation as Texas attempts to rebuild its infrastructure.

NUAI isn’t just selling a fast track to power; they are selling a permanent, structural hedge against grid inflation. For a hyperscaler looking to lock in margins on a $15 billion campus over a 20-year lease, that predictability is of paramount importance.

7. The Platform Potential

As a thought experiment, it is worth exploring why NUAI’s capital-lite merchant developer model might actually be more scalable than the asset-heavy path taken by players like Applied Digital (APLD). While APLD acts as a capital-intensive utility builder that must swallow massive equity dilution to fund its own projects, NUAI operates like an infrastructure General Partner (GP).

You can think of this like a traditional private equity firm’s structure. A PE firm carves out a piece of the equity pie for themselves even though they didn’t put down the upfront capital. Instead, they earn that equity through the effective management and execution of their limited partners’ (investors’) capital. NUAI is doing the exact same thing with dirt and megawatts. Rather than relying on easily copied intellectual property, they trade execution, physical site control, and experienced regulatory navigation for carried interest (a percentage of the total equity) in a BTM data center joint venture. Because NUAI isn’t tethered to its own balance sheet for billions in construction CapEx, they can move faster and manage multiple gigawatt-scale projects simultaneously.

While an asset owner has to finish one project before they can afford the next, a merchant developer with a repeatable playbook is only limited by how many sites they can lock down and how many execution partners they can call. In this scenario, NUAI isn’t just building a data center; they are building a franchise of high-margin, low-capital-intensity, highly predictable cash flows and immensely valuable equity stakes.

But this raises a critical bear argument. Critics rightly point out that the ERCOT rulebook is public, the SB 6 legislation is public, and pipeline maps are public. If the profit is this clear, what stops a massive private equity firm from copying the homework, or worse, why don’t the hyperscalers just do it themselves?

The defense against the private equity threat lies in the friction of the physical world. There is a massive difference between reading a regulatory framework and securing physical site control.

To get control of dirt, you have to win over the people who own it. It is worth asking why independent landmen still dominate the Permian Basin today. You would think supermajors like Exxon or Chevron with unlimited capital would just bring that entire function in-house to cut out the middlemen. They don’t, because a fourth-generation West Texas rancher sitting on a critical pipeline right-of-way usually doesn’t want to negotiate with a corporate lawyer from New York or Houston. They do deals with people they know, people who speak the localized language of the patch.

The exact same dynamic protects NUAI from a massive private equity fund trying to copy their homework. You have to find a landowner willing to sell, negotiate the surface rights, convince three different midstream operators to let you tap their lines, and navigate local county politics.

And frankly, that political friction is only going to get worse. Building gigawatt-scale data centers is becoming increasingly hostile at the local level. Communities are hypersensitive to the optics of tech giants draining local aquifers or driving up costs on their power grid. When a project inevitably hits a wall, whether it is a zoning dispute, a water rights challenge, or an unforeseen regulatory hurdle, a hyperscaler needs boots on the ground. They need an operator willing to take on the execution risk of cutting-edge, unconventional solutions to remove that friction. A mega-fund has the spreadsheets, but they don’t have the localized rolodex or the creative operational grit to push complex workarounds across the finish line. Acting as that professional problem solver is exactly what NUAI does best.

The answer to the hyperscaler question is the same reason a massive law firm rents a skyscraper rather than building it. Google and Meta are in the business of compute, chips, and models. They are not in the business of West Texas surface rights, gas mineral ownership, and localized electrical engineering disputes. They need powered land today, and they are willing to pay a massive premium to the developer who has already fought those localized battles to get it done.

Ultimately, the barrier to entry isn’t just buying dirt. It is the ability to assemble the exact soup of site-specific advantages, such as redundant pipeline intersections, existing high-voltage adjacency, and the legal workaround for the ERCOT queue. With a 230 GW backlog in Texas alone, the pie is simply too large for any single player to swallow. There will be many successful developers, but NUAI has differentiated itself by being the first to productize the BTM shortcut.

8. Hyper-Growth

It would be easy to look at the Ector County project and assume NUAI is a one-and-done story. It is tempting to write them off as just a lucky team that stumbled upon a golden ticket. But during the BofA call management made it excruciatingly clear that they are treating TCDC as the proof of concept for a much larger platform.

When asked about the roadmap for the balance of the year and into 2027, Charlie Nelson laid out the ambition clearly:

“We’re not a company that’s building a single asset. We’re a company that is building a platform, and that is the first asset to use that platform. And so, taking that platform, everything that we’re doing is designed for execution, managing hyper-growth, getting systems in place, continuing to expand the team, and then continuing to expand the asset base. Because our intention is not to stop at one or two assets, our intention is to get into hyper-growth mode and we’ve built ourselves to do that.”

This is the beauty of the GP model: because NUAI isn’t bogged down raising billions in equity to buy their own turbines and lay their own concrete, their only limit is their ability to execute the next deal. Nelson further confirmed that this isn’t just aspirational talk. He noted that they are already “in the midst of doing site selection for growth plans” and teased that there are “a lot more power sources, a lot of new stories, all sorts of fun stuff coming.”

If NUAI successfully lands a big four hyperscaler anchor tenant for TCDC the market will reprice them based on the NPV of those localized cash flows. And out of conservatism that is where I ended my model.

But the true 10x upside lies in what happens the day after that deal closes. What happens when they become the de facto problem solvers for any hyperscaler looking at West Texas?

They aren’t just selling a gigawatt of power. They have built the platform to sell the Permian Basin to Silicon Valley.

Subscribe now

Leave a comment

Share

Disclosure: I am long NUAI. I may buy, sell, or change my position at any time without notice. This article reflects my personal opinions and is for informational and educational purposes only. Nothing in this article should be construed as investment advice or a recommendation to buy or sell any security. I have not received compensation from NUAI, its management, or any affiliated party for this content. I have no business relationship with the company.

This update is essentially a play-by-play of an infrastructure heist. If you’ve been following my time-to-power thesis since January, you know the game: in a world where AI is bottlenecked by 7-year utility queues, the only thing that matters is which data center developer can plug in a hyperscaler’s compute first.

Last week, NUAI didn’t just move the needle; they effectively jumped the entire line. Between two major PRs and some fresh information from investor relations, the signal is clearly breaking in shareholders’ favor.

The 54-Acre Corridor (Major News #1)

The news from February 26th about the 54-acre land corridor is the most important piece of the puzzle. This isn’t just a land grab. This strip is the physical bridge connecting the TCDC site to a massive 1.5+ GW power nexus consisting of the Quail Run (550 MW) and Odessa-Ector (1 GW) power plants.

Before this purchase, NUAI was technically near power, but in the eyes of Texas regulators, they were worlds apart. In the Permian, “close” doesn’t give you the right to run pipes or wires across someone else’s property. By securing this specific corridor, NUAI has turned the TCDC campus into a physically integrated extension of the power plants’ footprint, legally qualifying as “adjacent.”

I’m convinced the anchor tenant made this a non-negotiable requirement. The acquisition was explicitly undertaken as part of “ongoing lease negotiations with a leading hyperscale tenant,” where securing the land “emerged as a key milestone in advancing discussions.” It allows the tenant to tap directly into gas, water, and electricity without begging for outside easements or dealing with the transmission and distribution (T&D) fees that can destroy project economics.

CEO Will Gray made the stakes clear:

“By taking the first step to secure this additional 54 acres, we are seeking to strengthen our strategic land position... and take a pivotal step toward finalizing a definitive lease agreement with our end tenant.”

The Regulatory Environment

Here is where the hidden math of Texas energy law comes in. NUAI can’t just sign a Power Purchase Agreement (PPA) to tap into that 1.5 GW nexus. Under the SB 6 regulations passed late last year (specifically PURA § 39.169), ERCOT is fiercely protective of existing capacity. A PPA legally classifies the data center as an “unaffiliated retail customer” asking to pull massive dispatchable power off the grid. ERCOT would flag it as a reliability threat. The PUCT might not issue a flat denial, but they would effectively kill the deal anyway by attaching economically fatal conditions—like hundreds of millions in transmission upgrade fees and strict, unmitigated curtailment mandates.

My thesis is that they are preparing for a total buyout. If NUAI buys the 550 MW Quail Run plant (estimated ~$550-$650M) to pair with the 54-acre corridor, their legal classification changes entirely. They go from being a heavily regulated retail customer to an integrated owner consuming their own behind-the-meter (BTM) generation on adjacent land.

But here is the catch: because NUAI missed the January 1, 2025 grandfather clause in SB 6, they still face a mandatory 180-day PUCT reliability review. Under the new law, the PUCT is legally required to mandate that the generator export its historical capacity back to the grid during an emergency. Without a backup plan, the moment a winter freeze hits, the state forces Quail Run to divert power to residential lights, and the data center goes completely dark—a risk no hyperscale tenant would ever accept.

This is why I am betting the hidden piece of the puzzle is a ~500 MW battery storage play. By investing ~$350M (post-tax incentives) into on-site storage, NUAI engineers a physical workaround to a legal mandate. When ERCOT hits the emergency button, NUAI can legally comply with the forced export order while instantly switching their servers to run on the battery reserves, turning a grid threat into a grid stabilizer without any downtime.

So why not just sign a PPA and build a battery? Because the buyout is what transforms the 180-day review from a regulatory roadblock into a technical negotiation, and that requires leverage. If NUAI were merely a PPA tenant, the PUCT could still swat them away or drown them in fees because a retail tenant has zero constitutional property rights to the generation asset. However, as the 100% owner of the power plant and the connecting 54-acre corridor, NUAI is shielded by the Texas Constitution’s “takings clause.”

By baking the battery shield into the initial plan, NUAI pre-solves the state’s biggest anxiety, the reliability gap, while simultaneously establishing the legal standing to argue that a flat rejection would constitute a multi-billion-dollar “regulatory taking.” The batteries aren’t just an equipment cost; they are the political and legal capital required to box the state in and buy a yes from PUCT. It transforms a 1 GW grid-drain into a self-balancing Private Use Network (PUN) that the state simply cannot afford (financially or legally) to block.

This isn't just a localized Texas loophole; it is rapidly becoming the blueprint for the entire digital infrastructure sector. The era of the hyperscale PPA is dying, choked out by grid delays and regulatory hostility. Look at Amazon's attempt to use a co-located PPA at the Susquehanna nuclear plant in Pennsylvania—it is currently bogged down in federal regulatory hell as grid operators try to attach crippling fees to the deal. The smart money has already pivoted to vertical integration. In December, Google dropped $4.75 billion to acquire Intersect Power specifically to build behind-the-meter energy parks paired with gigawatt-hour batteries. Two months later, TeraWulf bought the Morgantown power plant in Maryland outright to do the exact same thing. By buying Quail Run, NUAI would be executing the exact playbook the titans of the industry are using to physically and legally separate themselves from a failing grid.

But why would the current owner, Constellation/Calpine, actually agree to sell a cash-flowing asset during an AI boom? The incentive is rooted in a forced regulatory reality. Following Constellation’s (CEG) massive $26.6B acquisition of Calpine in January, the firm is in a aggressive deleveraging cycle. More importantly, regulators have already forced CEG to divest over 3,000 MW of capacity to settle antitrust concerns and prevent monopolistic dominance in ERCOT. Selling an older, 2007-era asset like Quail Run allows Constellation to satisfy these divestiture mandates and clean up its balance sheet, while offloading the complex SB 6 'reliability review' and battery requirements to a buyer with a hyperscale-backed balance sheet.

Still don’t believe me? NUAI has repeatedly stated they are targeting 1 GW for the TCDC facility. Tapping into the neighboring Quail Run plant would secure exactly 550 MW of capacity. 1 GW - 550 MW = 450 MW. Guess what other news came out last week? A deal for exactly 450 MW of gas engines for BTM power.

The Jet Engines (Major News #2)

On February 27th, they nearly doubled the BTM power agreement with Thunderhead from 250 MW to 450 MW. You don’t upsize by 200 MW on a whim. You do it because the hyperscaler on the other side of the table told you they need more power.

To pull this off, they’re using aeroderivative turbines through Turbine-X. These are basically modified jet engines. While every other developer is sitting on a three-year waitlist for industrial turbines, NUAI is tapping into shadow inventory—flight-certified cores that can be packaged and deployed in months, not years.

The kicker? NUAI isn’t even spending their own cash on this. Their partner, Thunderhead, is leveraging Harbert Infrastructure’s balance sheet to lock down the equipment. CEO Will Gray hammered this point home:

“Securing the critical-path power equipment for our anchor tenant's build-out significantly de-risks the project's supply chain and construction timeline—all without direct capital expenditure from New Era. This is our model in action."

That quote is the key to understanding how a micro-cap developer pulls off a massive infrastructure buyout without diluting shareholders to zero. Just like the turbines, the Quail Run acquisition and the 500 MW battery play will be financed entirely at the project level. The heavy lifting won’t fall on NUAI’s balance sheet; it will be underwritten largely by cash from their partnership with Primary Digital Infrastructure (PDI) alongside debt from private credit markets optimizing for WACC.

The Trump Administration Tailwind

The administration’s “Ratepayer Protection Pledge” is the perfect tailwind here. It’s a policy that effectively tells tech companies to bring their own power to every data center build from here on out. This transforms NUAI’s off-grid Permian strategy from a niche experiment into the federally preferred blueprint for fast-tracking AI.

IR Update

A fresh IR email is circulating on X that confirms they are in “advanced negotiations” with a tenant and are on track to meet the Q1 deadline. They’ve already started clearing the site and confirmed that soil testing came back positive.

Conclusion

My biggest worries for this thesis have been mitigated, and thus my official price target is now $20, pricing in an ~80% success rate in my model.

I don’t know of any stock more asymmetric right now than NUAI at $4.14.

Leave a comment

Subscribe now

Share

Disclosure: I am long NUAI. I may buy, sell, or change my position at any time without notice. This article reflects my personal opinions and is for informational and educational purposes only. Nothing in this article should be construed as investment advice or a recommendation to buy or sell any security. I have not received compensation from NUAI, its management, or any affiliated party for this content. I have no business relationship with the company.

• The US terrestrial power grid is permanently capping AI growth. Building gigawatt scale data centers on land is no longer an engineering problem. It is a political and regulatory impossibility constrained by five to seven year grid queues and a breaking water-energy nexus.

• Orbital data centers generate 45% more usable compute per watt. Terrestrial cooling taxes waste hundreds of megawatts per facility. Because of the physics of passive radiation in a vacuum, a 1 GW power budget in space yields 909 MW of actual compute compared to just 625 MW on Earth.

• Space is the only way the US can bypass China’s infrastructure advantage. The US cannot win a terrestrial infrastructure war against an authoritarian state that does not require building permits. Launching compute to orbit replaces decades of domestic regulatory gridlock with technological speed.

• Orbital compute holds an $18.2 billion Net Present Value advantage per gigawatt over terrestrial data centers. Building a 1 GW facility on Earth requires $26.6 billion in infrastructure capital expenditure and a six year grid interconnect queue. Deploying that identical capacity to orbit via 162 Starship launches costs just $11.6 billion and takes one year. This creates a $9.5 billion time to market premium and slashes operating energy costs from $0.075 to $0.0022 per kilowatt hour, easily absorbing a 4 percent hardware mortality rate.

Intro

For the last decade, scaling compute was a matter of landing an order for more GPUs and signing a utility contract. The bottleneck was silicon. Today, the bottleneck has shifted to the substation. The era of the easy, grid-connected data center is over.

Interconnect queue times and turbine backlogs now stretch five to seven years in Tier 1 markets. Nuclear Small Modular Reactors are frequently cited as the solution to this base load problem but are years away from commercial viability and face the exact same public pushback that stalled the existing grid: water.

Generating 1 MWh of electricity in the US evaporates between 300 and 700 gallons of water at the power plant. A 1 GW data center indirectly consumes over 10 million gallons of water every day just at the utility source.

Closed loop cooling is a thermodynamic compromise. Water conducts heat roughly 24 times better than air. Because of this physical reality, closed loop systems require massive fan arrays and higher compression ratios, which increases a facility’s total power consumption by 20% to 30%. For a 1 GW cluster, this cooling penalty draws an additional 200 to 300 MW. The extra electricity needed to run these fans forces the upstream power plant to work harder. That added power load evaporates an additional 2.4 to 3.6 million gallons of water every day off-site. You might save water on the data center property, but you still place a massive strain on the local watershed.

In regions like Hood County, Texas, a place historically known for its business-friendly environment, this resource bottleneck has triggered a political revolt. Just this month, residents demanded the resignation of county commissioners over a proposed 2,100-acre data center that threatened the local aquifer. Data centers are no longer viewed as economic development, but as direct threats to local wells and agricultural resources. County officials are realizing that greenlighting these massive water draws is a fast track to losing their seats.

US vs. China

AGI is a winner-take-all scenario. Whoever builds it first controls the platform that dictates global ideology. Historically, AI scaled through three levers: model parameters, dataset volume, and training compute. The first two have hit a structural ceiling. High-quality human text is exhausted, and intelligence gains have stalled from simply increasing model size.

This leaves compute as the only scaling law left to propel us to AGI. In a world of inference-time scaling, where models use massive reasoning loops to “think”, a nation’s collective intelligence is intresically tied to its total available power. A two-year lead in chip design is irrelevant if China can plug in an order of magnitude more silicon than the US.

If China deploys thousands of GW to inference while the US is capped at hundreds by regulatory gridlock, China effectively possesses a higher national IQ. They will solve fusion, crack encryption, and automate warfare while the US waits on substation permits.

China’s power infrastructure lead is staggering:

• By 2026, China’s installed power capacity hit 3.9 Terawatts (TW), triple the US capacity of 1.2 TW.

• China’s wind and solar capacity alone (~1.8 TW) is already larger than the entire US grid across all sources combined.

• In 2025, China added 543 GW of new power, outpacing US power infrastructure build-out by over 800%.

• Goldman Sachs predicts that by 2030, China is projected to have 400 GW of spare capacity. Conversely, the US faces a 44 GW shortfall by 2028.

The US cannot win a terrestrial infrastructure war against an authoritarian state. Under the “East Data, West Computing” initiative, China has nationalized its compute supply chain. They bypass domestic friction by state mandate—rerouting rivers for cooling and building dedicated power lines from coal hubs to GPU clusters.

The strategic stakes are absolute. If China reaches AGI first, seizing Taiwan and TSMC becomes an inevitability. By controlling the world’s most advanced chip supply while possessing a massive lead in power, China would gain permanent leverage over the West.

Data centers in space sound outlandish until you evaluate the alternatives. To counter China’s lead, the US must use Starship to build compute in orbit faster than China can build on land. This makes the AGI race a winnable game by playing to our unique strength: American innovation. The US finally leverages its one true trump card—a fully reusable SpaceX rocket twenty years in the making.

The Physics

Space is a fundamentally superior environment for high-density computing because it removes the friction of the atmosphere. On Earth, we are constantly fighting a density tax. As you pack thousands of GPUs closer together to make them faster, they generate so much heat that air cooling fails. This forces companies to build massive, expensive liquid cooling systems that eat up the power budget before it reaches a chip.

Terrestrial data centers operate at an average efficiency (PUE) of 1.6. In a 1 GW facility, 375 MW is lost just to keep the building from melting. You pay for 1 GW of electricity but only get 625 MW of actual work. In orbit, cooling is handled passively by radiating heat into the deep cold of space. This brings efficiency to a near-perfect 1.1. A 1 GW power budget in space yields 909 MW of work—a 45% increase in usable intelligence for the same energy input. To get that same 1 GW of work on Earth, a developer has to build a facility 60% larger just to pay the cooling tax.

In a sun-synchronous orbit, a satellite is bathed in constant, unfiltered sunlight with 40% more intensity than we get on the ground. This eliminates the need for massive batteries or a complicated power grid. While a subsidized data center in West Texas might pay five cents per kWh, space power costs roughly two-tenths of a cent.

Heat in a vacuum is removed via radiators. Because of the Stefan-Boltzmann Law, the power a radiator can dump scales to the fourth power of its temperature:

This creates a massive incentive to run hardware hot. In space, every extra degree you can push a chip makes the cooling system significantly more efficient. Running hot allows for smaller, lighter radiators, which cuts launch costs.

Finally, light travels roughly 50% faster through the vacuum of space than it does through the glass inside a fiber-optic cable. By using Starlink’s laser network, data can be routed around the globe faster than any terrestrial cable. At an altitude of 340 miles, the signal “drive” to an orbital GPU is roughly the same distance as a trip from Austin to the data centers in West Texas.

GPU Mortality

The trade-off for orbital efficiency is a jump in hardware mortality. On Earth, a Tier 1 data center is a sanctuary with stable temperatures and a thick atmospheric shield. Annual GPU failure rates there hover around 1%. In orbit, you pay a space tax in the form of radiation and mechanical stress.

Radiation flux in low Earth orbit is up to 1,000x higher than at sea level. While the slow accumulation of damage is manageable, individual high-energy particles are a constant threat. These particles cause bit flips where a zero becomes a one. If a 10,000 GPU cluster on Earth sees one error a day, that same cluster in space could see up to 40 errors every hour.

The consequences range from trivial to terminal. In the best case, error correction code fixes the flip in nanoseconds. In a middle case, the AI produces a typo and the load balancer simply routes the next request to a clean GPU. At worst, a particle triggers a latch-up, a short circuit that forces the satellite to reboot. This makes orbit much better for inference than training, as it is easier to lose a single request than a month of training progress.

While sun-synchronous orbits provide constant sunlight, satellites still face temperature swings of 5 to 10 degrees every 90 minutes from heat reflected by the Earth. Over a five-year mission, GPUs undergo roughly 29,000 of these thermal expansion cycles. This constant mechanical strain on solder joints, combined with radiation exposure, pushes annual hardware mortality to roughly 4%.

While it sounds nice, building a cluster out of radiation-hardened chips is not a solution. That silicon is typically three generations behind the state of the art, and TSMC will not re-tool entire factories for space chips while terrestrial demand is at a breaking point. The winning strategy is to use the best chips available and treat the cluster as a living machine. Each cluster requires a dedicated nanny chip: a hardened controller that monitors thousands of sensors to manage survival.

The Black Swans

The physical environment of low Earth orbit presents existential risks that mirror the internal mortality of the chips. While a nanny chip can manage a bit flip or a coolant leak, it cannot protect against systemic orbital collapse.

SpaceX has applied for one million satellites, a 67x increase over the current population. Because collision risk scales exponentially, this represents a 4500x increase in the probability of impact, requiring tens of millions of automated dodges per year. Just as a single high-energy particle can kill a GPU, a single solar flare that knocks out telemetry for 24 hours could trigger a Kessler chain reaction.

Furthermore, the constant de-orbiting and burning of hardware to manage the 4 percent mortality rate deposits unknown quantities of alumina particles into the ozone layer.

The Economics

The financial argument for orbital compute is a massive infrastructure arbitrage. A 1 GW data center on Earth is burdened by an atmospheric tax that forces developers to build 1.6 GW of power just to cover cooling losses. Between land acquisition, grid connection fees, and industrial HVAC plants, the capital expenditure for a terrestrial cluster has climbed to $26.6 billion. In contrast, the space-based model pencils out at $11.6 billion, including the cost of Starship launches at $200 per kg. This creates a $15 billion upfront discount before a single watt of power is consumed.

However, the launch itself is strictly governed by physical volume, not just weight. To hit that $11.6 billion CapEx, the math demands packing 75 satellites into every Starship. If developers rely on traditional, rigid solar panels and bulky radiator frames, they will run out of space inside the payload bay long before hitting the rocket’s mass limit, capping out at roughly 20 satellites per launch. That volume constraint would balloon a 1 GW project from 162 launches to over 600, instantly wiping out the financial advantage and the speed-to-market premium.

To bypass this, the industry is forced to invent a new class of flexible architecture—a unified compute wing. Whoever solves the materials science of integrating rollable liquid cooling loops directly into deployable solar blankets holds the key to making the packing density work.

Once in orbit, managing the 4% annual mortality rate uncovers a severe environmental bottleneck. A 4% failure in a 1 GW cluster represents $2 billion in lost revenue, but without in-space repairs, maintaining a massive constellation means de-orbiting and burning up dead satellites in the atmosphere every year. Vaporizing that much aluminum deposits unknown quantities of alumina particles directly into the ozone layer, threatening a systemic environmental crisis.

SpaceX can solve both the financial and environmental risks with a dedicated maintenance architecture. A Starship “Chomper” variant can ingest these satellites into its cargo bay, where Optimus robots perform high-precision hardware swaps. At a mission cost of roughly $450 million, recovering $2 billion in revenue yields a 440% return on investment, while simultaneously keeping dead hardware out of the upper atmosphere. This capability would also allow SpaceX to hot-swap older chips for next-generation silicon, upgrading the intelligence of the constellation without the cost of a full redeployment.

Even if the $9.5 billion time-to-market credit and the Optimus replacement capabilities are ignored for the sake of conservatism, the initial construction gap makes orbit the superior economic choice. For the model to hold, we only need to believe in a few core variables: Starship reaching $200 per kg, the successful engineering of high-density flexible cooling wings, the continued gridlock of the terrestrial power queue, and the ability of autonomous nanny chips to manage hardware failure.

This is how the US wins the race to AGI.

View the full space data center model: Here

Leave a comment

Share

Disclosure: This article reflects my personal opinions and is for informational and educational purposes only. Nothing in this article should be construed as investment advice or a recommendation to buy or sell any security.

Key Takeaways

• AI has reintroduced linear Cost of Goods Sold (COGS) to software. The AI Inference Tax will structurally compress gross margins from ~95% to ~70%.

• Valuation follows a step function, not a gentle slope. A drop in gross margin from 95% to 70% looks like a minor haircut, but it mathematically triggers a ~65% collapse in intrinsic value by destroying the leverage that made software worth 30x revenue.

• Cheaper inference tokens will not save software margins. Jevons Paradox ensures that efficiency gains lead to exploding consumption as competitors reinvest every saving into deeper reasoning and always-on agents.

• Usage-based billing destroys the predictability that made SaaS multiples so rich. Consumption behaves like a cyclical utility leading to worse quality of earnings and lower valuation multiples.

• The software endgame is a liquidation play, not a growth story. Firms like Bending Spoons will buy legacy assets at a discount and use agents to gut the old SaaS org chart for pure cash extraction.

Intro

For the last decade, being a SaaS founder was the ultimate arbitrage. You cooked the meal once and served it a billion times. The marginal cost for the ten-thousandth user was effectively zero. It was the economic equivalent of Jesus feeding the 5,000 with a couple of loaves and fishes.

This economic miracle is why the market awarded these companies 20x and 30x revenue multiples. You weren’t just buying growth. You were buying the Gold Standard business model: subscription-based software.

SaaS was the ultimate compounder. It offered predictable, recurring, and gravity-defying economies of scale. Analysts modeled cash flows 20 years out, assuming that 95% GRR and 110% NRR were laws of physics. These companies weren’t treated like risky equities; they were treated like government bonds, where revenue retention was considered as guaranteed as the US Dollar itself.

The model was so bulletproof that even Private Equity, notorious for demanding durability, felt comfortable piling massive leverage onto these businesses. The subscription revenue was supposed to be a sure thing.

But the menu has changed. We are witnessing the McDonaldization of software. This is a fundamental shift from high-margin luxury to a high-volume, thin-margin, variable-cost utility.

This forces a pricing model shift from predictable seat-based subscriptions to metered, usage-based billing. That move protects gross margin, but it destroys the thing SaaS sold investors: recurring, forecastable revenue. Consumption behaves like a utility. It fluctuates with budgets, workloads, and throttling. The result is worse quality of earnings, with less visibility, more volatility, and lower valuation multiples.

Here is why the quality of earnings in SaaS has permanently degraded.

The Death of the Asset-Light Model

Traditional SaaS enjoyed upwards of 95% gross margins because software scaled sub-linearly. But AI has reintroduced COGS to a sector that had forgotten they existed. In the pre-AI era, the only friction on a software P&L was the AWS Tax. These were the basic hosting fees required to keep the lights on. Crucially, these were step function costs rather than linear ones. You paid up front to spin up the infrastructure, but the marginal cost of adding the 10,000th user was effectively zero. Because of this, gross margins of 95% were the norm. The revenue scaled exponentially while the costs stayed relatively flat.

Fast forward to present day and take a look at companies like Duolingo for example. In late 2025, despite crushing revenue estimates with 41% YoY growth, their gross margins compressed to 72.5%1. The reason is simple. Every time a user asks for a grammar breakdown, tokens are burned. Every time they video chat with Lily, a GPU hums. It consumes real power to manufacture the hundreds of tokens required for that single interaction. These costs bleed directly into the P&L. In their Q3 2025 letter, management explicitly cited increases in generative AI and hosting costs as the driver for this compression.

Traditionally in SaaS gross margins expand with scale. Every incremental dollar of revenue should be accretive, driving margins higher. But with Duolingo, we are seeing the exact opposite. If a hyper-growth compounder growing at 40% annually cannot outrun the inflation of AI inference costs, nobody can. Duolingo is the canary in the coal mine as they are on the bleeding edge of AI implementation. On a leaked OpenAI enterprise customer list from late 2025, Duolingo was ranked number one. Once the rest of the slower-moving SaaS index is forced to execute similar rollouts just to stay competitive, this margin collapse will become universal.

To fight this, the largest software players are attempting to back their way into better margins by building their own inference clusters. They realize that renting from providers is a slow death, so they are trying to become mini-AWS instances. This creates a new structural nightmare. By moving from renting tokens to owning silicon, once asset-light companies are suddenly saddled with massive hardware depreciation, specialized engineering overhead, and exploding energy costs. They are trading variable token costs for permanent, massive Capex. Whether you rent the GPU or buy it, the Inference Tax caps your margins at ~70%.

But the expense side is only half the nightmare. The revenue engine is simultaneously seizing up. The SaaS terminal value figures were built on the assumption that heads in beds would grow indefinitely. Per-seat pricing was a proxy for human labor. But in a world where a single AI agent can automate the output of ten junior analysts, that proxy is broken. Data from late 2025 shows that AI agents complete complex digital tasks 88% faster than human workers.2 If your customer replaces a 100-person department with 10 human supervisors and a fleet of agents, your revenue base and TAM both face a structural 90% contraction. Legacy vendors are now trapped. Stay on per-seat billing and watch revenue evaporate, or pivot to usage-based pricing and watch valuation multiples implode as the market stops treating software like a predictable annuity and starts treating it like a cyclical utility.

This headcount collapse exposes a dark reality about the legacy software model. For two decades, vendors were financially incentivized to build friction into their products. If a platform required manual data entry or complex navigation, the customer had to hire more people to manage the workload. Because the vendor sold by the seat, inefficiency was a direct revenue driver. The industry fed on license bloat for years, intentionally capitalizing on corporate bureaucracy and operational waste. Enterprises ended up with thousands of underutilized licenses simply to overcome the friction of the software itself.

AI agents are now popping this bubble. An agent does not care about a user interface or a multi-step workflow. It interacts with the software programmatically at the API layer. The friction that used to sell seats is now a technical bottleneck that prevents a customer from scaling their agentic fleet. As customers move toward agentic workflows, they will stop paying for the cost of human inefficiency. They will move toward whichever provider allows for the highest volume of work with the least amount of human oversight. The era of capitalizing on unnecessary complexity and deliberately bloated software is over.

This transition also breaks the traditional Go-To-Market motion. In the old world, you sold a vision to a CEO based on projected headcount. In the new world, you are selling a utility API call that plugs into a customer automated workflow. You cannot justify a massive sales commission for a usage-based product that starts at zero. When you move from selling seats to selling APIs, you move from the strategic board room to a line-item in a developer budget. The moment a competitor offers a lower price per token, the customer swaps the API key. There is no user friction to stop them because there are no users, only agents.

This is the new reality. You are not selling a zero-marginal-cost license anymore. You are running a digital assembly line where every single interaction has a price tag and every relationship is a commodity. The software purchases a raw token, wraps it in value using personalized user data or a proprietary moat, and spits out an interaction. Assembly lines do not trade at 30x revenue.

Despite massive 41% YoY revenue growth, Duolingo’s gross margins are falling. This proves that new AI-heavy users are dilutive to profitability. While the 30% App Store Tax always made 95% margins impossible, the new AI costs are compressing them even further.

The First Principles of Valuation Gravity

If you don’t believe the narrative, look at the math.

Businesses are valued on the sum of the free cash flow they produce over time, discounted back to the present day. Over the long term, stock prices track free cash flow per share. They do not track revenue.

To estimate the free cash flow ceiling for a business, you look at gross margins. Revenue minus the cost of goods sold is the gross profit. This is the fundamental gate that determines what can actually drop down to the bottom line for the most theoretically perfect operator imaginable.

At 95% margins, you have massive leverage. Every new dollar is 95 cents of dry powder. Once fixed costs are covered, practically every incremental dollar is pure profit. At 70%, that leverage breaks. You need 36% more revenue just to generate the same gross profit dollars as your high-margin peer.

The compression isn’t linear; it is geometric. If gross margins fall from 95% to 70%, your free cash flow margin likely collapses from 45% to 20% because your operating costs like SG&A and R&D are largely fixed. That isn’t a minor valuation haircut. That is a 70% wipeout of the intrinsic value of the business.

This 70% figure is actually a conservative baseline. It only accounts for the immediate destruction of current cash flows and does not factor in the inevitable expansion of the discount rate. When a business model shifts from a predictable software annuity to a volatile, high-CapEx structure with usage-based revenue streams, the fundamental risk profile changes. Even a minor upward adjustment in the discount rate to reflect the degradation of confidence in future predictability sends the terminal value into a free fall.

Subscribe now

For a B2C AI app, over 50% of every revenue dollar is eaten by the App Store and AI compute costs before paying a single employee. For non-App Store SaaS, these same compute requirements are dragging margins from 95% down to ~72%3. As the “Inference Tax” rises, the structural value of the business collapses.

The Efficiency Trap (Jevons Paradox)

The most common rebuttal to this thesis is that token costs are collapsing. In theory, as tokens get cheaper, margins should snap back to 95%, right?

That sounds reasonable until you remember Jevons Paradox. When something gets cheaper, we don’t conserve it. We consume more of it. LEDs didn’t reduce lighting; they lit up every parking lot, sidewalk, and kitchen cabinet in America. Faster internet didn’t cut bandwidth bills; it gave us autoplay, 4K, and a video feed that never ends. Storage didn’t make data lean; it made us stop deleting anything and start filming our entire lives. Compute is no different.

Yes, the cost per token is falling. That fact is real. It is also meaningless on its own, because the bill is not set by price alone. The only equation that matters is:

Compute spend = (price per token) × (tokens per task) × (tasks per user) × (active users)

If price falls 10x but tokens per user rises 10x, nothing gets cheaper. The bill stays the same. And in practice, it rises, because competition forces you to reinvest every efficiency gain into a better product, not a cheaper one.

That’s the real driver. As soon as tokens get cheaper, teams don’t hold the product constant and celebrate margin expansion. They expand the product itself. They shove more context into the prompt, run more reasoning in the background, add more verification steps, and push toward more autonomy. The market rewards better UX, not lower GPU bills, so the savings don’t drop to the bottom line. They get spent.

Token consumption climbs in three phases:

1) Standard retrieval: This is the basic chatbot era. Short prompt, short answer. A few hundred to a few thousand tokens and you move on.

2) Reasoning models: Now the cost moves into the dark. The interface shows one clean paragraph, but the model is generating far more than you can see. OpenAI counts reasoning tokens because they are real tokens. Epoch AI has shown reasoning models output around 8x more tokens than non-reasoning models on benchmarks. That is the trade. You buy fewer mistakes by burning more tokens.

3) Always-on agents: A chatbot is a transaction. An agent is a machine left running. It watches your inbox. It reads the entire thread. It pulls the doc. It checks the calendar. It queries the CRM. It drafts the email. It follows up. It checks its own work. It loops.

And loops are token-hungry by design.

One documented computer-use agent run consumed 275,219 input tokens in a single session. That is one task, one agent, one run. Scale that across thousands of employees, or millions of consumers, running in the background all day.

Even the tooling ecosystem is admitting this out loud. Anthropic tells builders to cap tool outputs because agents can chew through context so fast. Their default cap for Claude Code tool responses is 25,000 tokens. That is not an edge case. That is the workflow being poured into the context window.

Cheaper tokens simply force higher token volumes.

There is no AI Lite equilibrium. If a competitor gives you deeper reasoning, longer memory, always-on assistants, and multimodal help for the same price, you match it or you get labeled obsolete. So every efficiency gain becomes more features, more context, more agent steps, more tokens.

Token prices are falling. Token requirements are rising faster.

While the cost per token has collapsed (purple line), the explosion in token volume required for modern “reasoning” models (red line) keeps the total bill high.4

The Edge Illusion

The edge thesis is the comforting conclusion everyone wants. Eventually the models are good enough, inference runs locally, and the Inference Tax disappears. The problem is that good enough is not decided by your CFO. It is decided by your competitor.

Edge will absolutely happen. A meaningful share of AI will move on-device because it is faster, cheaper, and private. Low-stakes, high-frequency tasks will live there: autocomplete, content suggestions, lightweight summarization, transcription, simple copilots. Offices will buy AI PCs. Some large enterprises will rebuild on-prem GPU rooms for sensitive data or heavy internal usage. That will shift some compute cost off the vendor’s P&L in certain cases.

But this is where people make the wrong leap. Offloading compute does not restore 95% SaaS economics because gross margin was never the whole story. The old SaaS miracle was not just high margins. It was that the vendor owned the full stack and could deliver one uniform experience to every user with near-zero incremental friction. The moment inference becomes something the customer must provision, maintain, and power, you have changed the product. You have turned software back into infrastructure.

That shift blocks a return to the old model in two ways. First, the vendor loses pricing power. If the customer is buying GPUs, paying the electric bill, and living with the refresh cycle, they will not keep paying a premium subscription as if the vendor is providing the intelligence for free. They will demand lower pricing or they will switch to a cheaper wrapper, because the expensive part is now customer owned and increasingly standardized. If the customer is paying for the GPUs, the vendor cannot price the subscription as if it is delivering the compute. The surplus shifts to whoever owns the infrastructure. Second, the experience fragments. Hardware varies, updates lag, IT policies differ, and performance becomes a function of what machine you happen to be on. SaaS won the first time because it removed this exact problem. Centralization gave you consistency. Edge reintroduces variability. On-prem does not eliminate cost. It reintroduces IT labor, downtime risk, and upgrade friction that SaaS spent a decade eliminating.

The only way edge would fully solve the cost problem is if companies could freeze capability to protect margin and still retain customers. In practice, that means shipping a weaker product and hoping the market does not punish you for it. History is clear about how that ends. Carnegie did not win steel by limiting investment. He reinvested until everyone else looked slow and expensive. Kodak had digital technology and still tried to defend film margins. The market did not reward caution. It punished hesitation. Blockbuster defended store economics while Netflix built streaming. BlackBerry defended its hardware model while Apple reset expectations. In each case, the incumbent did not fail because it was incompetent. It failed because it treated margin preservation as a strategy while a competitor treated product improvement as leverage.

Mobile makes the edge story even harder. Modern SaaS is mobile by default, and mobile runs on batteries. Even if Apple pushes more inference onto the iPhone, it will be great for small tasks and bad for heavy ones. Long context reasoning, multi-step agents, and sustained workloads drain battery, generate heat, and throttle performance. A phone running hard inference is a phone that dies before lunch, which makes the user experience worse, not better. And the hardware rollout problem is unavoidable. Consumers keep phones for years. Enterprises roll devices on multi-year cycles. If your product’s intelligence depends on the newest hardware, you create a two-tier experience across your user base, which becomes a support nightmare and a retention risk.

That is why running older models locally is not a stable equilibrium. The model is not competing against last year’s model. It is competing against the best experience available today. If a competitor offers deeper reasoning, longer memory, background agents, and multimodal workflows at roughly the same price, you cannot respond by choosing the cheaper model and expect retention to hold. You match the experience or you lose users. Matching the experience requires compute somewhere in the system.

There is a second problem that is even more dangerous. Suppose you succeed in pushing inference on-prem, freeze capability, and return to 95% gross margins. You have not made the business safer. You have made it more vulnerable. High gross margins are an invitation. They signal slack. They create room for someone else to eat the Inference Tax and still build a profitable business.

If you are operating at 95% gross margins by avoiding heavy inference, you have created a gap between what you charge and what it costs to deliver the modern experience. A competitor can step in, take the hit on compute, run at 40% or 50% gross margins, and still be profitable. They do not need your margin profile. They need your customers. The higher your margins, the more room you leave for a new entrant to price below you, ship a better product, and still make money.

Even in a world full of AI PCs and on-prem GPU rooms, the most valuable workloads remain heavy. Long context windows, tool orchestration, verification passes, shared memory, and cross-team coordination demand real infrastructure. Some inference may run locally, but the expensive portion does not disappear. It becomes hybrid. And when compute shifts on-prem, it remains a real cost. The burden shows up in hardware purchases, depreciation, power consumption, cooling, and refresh cycles. The marginal user still requires marginal infrastructure.

Cheaper compute does not end the arms race. It accelerates it. Edge changes where the cost sits and who bears it. It does not restore the economic miracle of zero-marginal-cost software.

The Collapse of Revenue Predictability

This is the most dangerous inversion. Historically, SaaS was valued on predictability and engagement. The more a user loved your product, the better.

Today, that logic breaks. If you stay on a flat-rate subscription, your power users become your biggest financial liability. They drive the most inference, burn the most tokens, and quietly compress your margins.

So you pivot to usage-based pricing to protect the P&L. You might bill per token or per workflow run. Or you might price on outcomes like tickets resolved, calls handled, documents processed, or deals sourced. The label changes, but the economics do not. You have moved from a contract to a meter.

Seat-based subscriptions behave like annuities. They renew on a calendar, expand with headcount, and rarely shrink unless the customer churns outright. That structure is what let investors underwrite revenue years out.

Usage revenue does not work that way. It moves with workloads, budget scrutiny, and optimization. Customers cap spend. Finance teams audit consumption. Engineers tune prompts and workflows to drive the bill down. In a slowdown, usage gets cut immediately. There is no renewal moment to defend, and revenue can fall without a single logo leaving. Even positive behavior becomes a forecasting problem, because the largest customers have the strongest incentives to throttle, optimize, and renegotiate once spend is tied to activity.

That shift collapses visibility. When visibility collapses, the asset stops being treated like a long-duration bond and starts being priced like a cyclical utility. The discount rate rises, confidence in future cash flows falls, and the premium multiple disappears.

The Macrohard Threat

The final nail in the coffin is competition.

Incumbent SaaS companies carry a heavy cost stack. Sales and support orgs, HR, management layers, and the overhead of a business built for the subscription era. Now add inference on top. Pricing has to cover both the legacy machine and the compute bill.

xAI is aiming directly at that structure with Macrohard. In the February 2026 xAI all-hands presentation, Musk described Macrohard as “intended to do full digital emulation of entire companies,” and later called it “emulation of entire human companies,” focused on businesses “where the output is digital.” In plain terms, the goal is not just a better tool. It is a system that can do the work a company does on a computer, end to end.

That is what changes the economics. If a new entrant can deliver comparable outcomes with far fewer people, its cost base is structurally lower. It can price below incumbents, move faster, and still earn acceptable margins. Even if headcount never goes to literal zero, the direction matters. Competition shifts toward unit cost and execution speed, and pricing pressure spreads across the category.

Macrohard is also not framed as a side project. Musk said it “over time actually will probably be our most important project,” because so much of the economy is made up of companies whose outputs are digital. If that kind of automation becomes credible, premium ARPU becomes harder to defend. Customers stop paying for overhead. They pay for results, and results get benchmarked against the lowest-cost operator that works.

So Who Survives?

Assume the hard version of the future. AI takes most tasks that can be verified cheaply and objectively. If an output can be generated, checked, corrected, and repeated at machine speed, the workflow collapses into automation. What used to be a standalone product becomes a feature inside a broader system.

Even the physical world is not a permanent moat. If agents gain a physical presence through humanoid robots, many structured, repeatable tasks become automatable as well. The constraint shifts. It is no longer about whether something happens on a screen or on a factory floor. It is about whether the outcome can be verified cheaply and looped at scale.

The durable survivors cluster where verification is not the whole story. They sit at the intersection of accountability, rights, and trust.

Regulation and Systems of Record

In regulated environments, “mostly right” is unusable. Someone must certify the filing, approve the transaction, sign the report, and carry liability if it is wrong. The value lives in audit trails, controls, permissions, validation logic, and legal defensibility. Even if agents perform the underlying work, the system that regulators recognize as authoritative still has to exist. That layer retains pricing power because accountability cannot be outsourced to a probabilistic model.

Identity, Security, and Governance

As autonomous systems gain access to more workflows, control layers become more critical. Identity, permissions, policy enforcement, logging, provenance, and monitoring are not optional. When an automated system misfires, organizations must prove what happened and who had authority. That requirement intensifies as automation scales.

Financial Rails and Risk

Money, settlement, reconciliation, tax, payroll, and compliance remain governed by counterparties and regulators. The workflows will automate. The liability does not disappear. The infrastructure that enforces rules, tracks exposure, and produces defensible records remains essential.

Rule of thumb: if a task can be verified cheaply, it gets automated and competed down, whether digital or physical. What holds value is not content generation or labor. It is verification authority, control, and accountability under real-world constraints.

The Private Equity “Agent LBO”

This sets up a contrarian outcome: the biggest winners from the repricing of SaaS may be the buyers, not the builders. When growth stalls and multiples compress, a lot of software stops trading like a compounder and starts trading like an operating asset. The product might still be useful. The customer base might still be sticky. But the market will no longer pay up for the old promise of effortless expansion.

That creates the opening for an agent-style LBO, and we do not have to wait for 2027 to see the playbook. Bending Spoons is already running it in public.

Their approach is simple and repeatable. They buy mature software brands with large installed bases, then they rebuild the cost structure around the reality that software is no longer scarce. They centralize operations onto a common internal platform, remove layers of headcount that were built for the growth era, and push pricing. If users churn, fine. The goal is not to win the market. The goal is to maximize cash flow from the users who remain.

This is the core insight. In a world where the marginal cost of “doing work on a computer” keeps falling, the old SaaS org chart becomes the liability. Sales teams, support orgs, marketing layers, and internal process sprawl were tolerable when the market paid 20x to 30x revenue for predictable subscription growth. When that multiple collapses, the company has to be restructured to match its new reality.

The agent LBO is what happens when you take that logic to its conclusion:

• Buy a stalled software asset at a low multiple

• Strip out the legacy cost stack built for the subscription era

• Replace as much repetitive internal work as possible with automation and agents

• Run the business for cash, not for growth

Bending Spoons is effectively proving that for a large class of software, the value is not in building new features forever. The value is in owning distribution, owning the installed base, and operating the product with a radically lower cost structure than incumbents assume is possible. If this model spreads, it will not just reshape software margins. It will reshape who captures them.

The Great Repricing

Let’s be clear: The software isn’t going away. But the financial asset class we called SaaS is dead.

For a decade, investors didn’t just buy technology; they bought certainty. They bought the mathematical inevitability of 110% NRR. They treated code like a government bond that just happened to be printed by Silicon Valley.

That bond has matured. AI has stripped away the scarcity that protected software margins. When intelligence becomes a commodity and switching costs drop to zero, the annuity model collapses into a war of attrition.

The value in this ecosystem is not disappearing. It is migrating. It is fleeing the application layer and settling in the physical layer. It is moving from the code to the current. The margin that used to belong to the software founder now belongs to the energy provider, the data center REIT, and the chip manufacturer.

We are witnessing the end of the zero-marginal-cost miracle. The magic of SaaS has been replaced by the physics of manufacturing. You are no longer running a high-leverage software monopoly; you are running a digital factory. And factories do not trade at 30x revenue.

Subscribe now

Leave a comment

“For very soon, maybe even later this year, we'll be producing more chips than we can turn on... The fundamental bottleneck is electricity.” - Elon Musk

January 2026

“The biggest issue we’re having now isn’t chips — it’s power. If you can’t build close to power, you’ll have a bunch of chips sitting in inventory you can’t even plug in. It’s not a chip supply problem anymore — it’s a power problem.” - Satya Nadella

November 2025

“Every single data center in the future will be power-limited. Your revenues will associate with the power you have to work with.” - Jenson Huang

January 2025

“The abundance of AI will be limited by the abundance of energy... The cost of intelligence will converge to the cost of energy.” - Sam Altman

October 2025

“I do not believe there is an AI bubble... What investors are really focused on is the long-term capacity for AI... Without a sustained push into AI infrastructure, the sector could face bottlenecks that stifle growth." - Larry Fink

January 2026

“Overall in the industry, maybe the bottleneck is power... We’ve brought in 3.8 gigawatts of capacity in the last year... and expect to double our overall capacity by the end of 2027.” - Andy Jassy

November 2025

Share Price: ~$7.60
Market Cap: ~$400M
Shares: ~53.4M basic, ~62.8M fully diluted today
Insiders: ~19.2% ownership (CEO ~1.65% ownership; 880,000 shares)

Summary:

I believe companies are rewarded in proportion to the difficulty and scale of the problems they solve. NUAI is going after arguably the biggest constraint in AI right now: time to power.

Across the AI landscape, the bottleneck has shifted. The biggest names have been pushing the same narrative: power and interconnect timelines, not GPUs, is the limiting reactant. Grid interconnect backlogs for hyperscale data centers are now measured in years, not months. If you want to compress time to power from years to months, you need a strategy that does not depend on a slow-moving and bureaucratic utility queue.

Behind the meter (BTM) power is the only way to bypass these interconnect timelines. Nuclear is compelling long term, but SMRs are not solving 2026 capacity. Fuel cells are interesting but expensive at the 1GW scale. Wind and solar help at the margin, but hyperscale AI loads require a firm baseload that runs 24/7. For near-term delivery at scale, natural gas generation is the most deployable option at GW scale.

NUAI’s roots are in natural gas, specifically the Permian Basin, which is one of the most important gas regions in the U.S. and has structural pricing dynamics (including WAHA) that can create very low marginal fuel costs at the right locations. NUAI is positioned to combine local energy know-how with a large, strategically located site to deliver fast, abundant, low-cost, reliable power to AI data center customers.

NUAI’s flagship project, Texas Critical Data Centers (TCDC) in the Permian, is designed to support a 1GW+ campus with a timeline that is competitive specifically on time to power. In an environment where hyperscalers are optimizing for speed and certainty, that is a major advantage. Management has stated they expect to announce an anchor hyperscaler tenant by the end of Q1 2026. If they execute, NUAI is one contract away from being valued like a real AI infrastructure development platform (think APLD) instead of a micro-cap.

Finally, regulation is a tailwind. The Trump administration is signaling that hyperscalers cannot lean on the grid and push costs onto normal Americans. Translation: bring your own power. Texas is reinforcing that with SB6 and ERCOT’s large-load rules. NUAI is positioned exactly for that outcome.

Company Description:

NUAI is building powered land and powered shell AI infrastructure in West Texas, centered on its flagship Texas Critical Data Centers (TCDC) site, about 438 acres outside Odessa in the Permian Basin. The plan is a 1GW+ campus built in phases, with the core advantage being behind-the-meter natural gas generation.

NUAI’s economic model is project-level. They plan to monetize TCDC primarily through long-term triple-net (NNN) leases to a hyperscaler anchor tenant and additional AI tenants. On top of that, NUAI intends to retain equity ownership at the project level, so as the campus gets financed and built out, NUAI participates in the residual value of the asset, not just a one-time land sale or a development fee. NUAI’s upside is a mix of durable lease cash flows and project equity value once the anchor tenant and financing package are locked in.

Market Opportunity:

Every enterprise’s productivity increasingly maps to how much compute you can reliably access. The market data reflects that: McKinsey’s forecasts global data center demand growing from ~55 GW in 2023 to ~219 GW by 2030 (about 4x), while Dell’Oro’s forecast shows annual global data center IT capex rising from ~$349B (2024) to ~$800B (2028). NUAI is positioned to capitalize on one of the largest infrastructure buildouts in human history.

Share Structure:

NUAI has about 53.4M shares outstanding and about 62.8M fully diluted today. For valuation, I use a more conservative ~71.5M diluted share count by 2029 to account for the dilution required to fund development and clear existing warrants.

The obvious question is how a ~$400M micro-cap funds a multi-billion-dollar campus. The answer is project financing, not corporate financing. Once an anchor tenant is signed, TCDC can be capitalized with a mix of non-recourse project debt and infrastructure equity, with NUAI retaining project-level equity and contributing land and development work. Insider ownership is roughly 19.2%, which is meaningful alignment at this stage. The CEO owns about ~1.65%.

NUAI’s partnership with PDI (announced 1/16/26) materially improves the financing probability here. PDI was founded by leaders who have already built at hyperscale, including Bill Stein (former CEO of Digital Realty) and Dave Ferdman (co-founder and former CEO of CyrusOne), alongside experienced digital infrastructure investors and developers. They have been involved in large AI data center development work, including the Abilene project tied to Crusoe and Blue Owl, which is the type of capital stack NUAI ultimately needs once an anchor tenant is signed. The partnership does not guarantee financing, but it improves credibility with tenants and capital, and significantly increases the odds that TCDC is completed.

Management and Board:

CEO Will Gray is a Midland native who grew up living and breathing the Permian. That local credibility matters in practice for permitting, county-level relationships, and moving a project through local stakeholders without being viewed as an outsider. He comes from an operating background in energy, and the board has real infrastructure and capital markets experience. P.J. Lee (EverStream Energy Capital) and Trent Yang (Galway Sustainable Capital) have backgrounds in financing and scaling infrastructure platforms. The job is to secure power, structure project-level capital, close long-term hyperscaler deals, and execute a large build on schedule, by bringing in best-in-class partners instead of trying to do everything in-house. NUAI’s role is to assemble the site, coordinate the stack, and retain project-level equity while specialists execute each layer.

Competition:

The closest public comp is Applied Digital (APLD). It’s a good reference point because the market has already shown how it prices scaled AI powered land and powered shell developers. The key difference is the power strategy. APLD is largely a grid and utility-interconnect story. NUAI is explicitly behind-the-meter and power-first in the Permian, optimized to compress time to power.

Beyond APLD, the competition is hyperscalers self-building, private infrastructure platforms, and large data center developers moving into AI campuses. But these deals are won on one variable: who can deliver first. NUAI does not need to beat everyone. They need to be one of the few who can put credible power and a credible timeline in front of a tenant.

Financials & Valuation:

At this stage, NUAI’s value is driven by TCDC’s economics and the probability of execution, not reported earnings. The way I look at it is simple: what does TCDC produce at scale, what is the path to get there, and what is that worth today after probability-weighting and dilution. I use two lenses: (i) a project NOI / cap-rate model, and (ii) an APLD MW comp as a check on how public markets price capacity once it is real.

My probability-weighted NOI / cap-rate model implies ~$1.18B of equity value, or about ~$16.49 per share. My success case (60% probability) implies ~$24.20 per share based on full execution. My failure case (40% probability) implies ~$4.90 per share based on limited execution.

As a secondary check, the APLD MW comp implies an ~$4.26B market cap for NUAI, or about ~$69.09 per share, assuming an APLD market cap of ~$10.1B as of Jan 26, 2026. I do not underwrite to the APLD comp, but it is a useful check on how public markets could re-rate these stories once a tenant is signed and the financing path is credible.

My valuation covers TCDC only. Management’s goal is to repeat this playbook and build a portfolio of powered campuses over time, but I am not giving credit for that optionality in the numbers above.

NOI / Cap Rate Model:

APLD Comp Model

View the full financial model: Here

Catalysts:

Over the next 12–24 months, the key milestones are the anchor hyperscale tenant announcement (the CEO has said this is expected before the end of Q1 2026), followed by FID, project financing (Blue Owl-type capital), and early construction and procurement milestones that prove the timeline is credible.

Financing Risk:

This is a multi-billion dollar project, so the financing stack is the gating item. The clean path is a long-term lease with a creditworthy tenant that supports non-recourse project debt plus infrastructure equity, with NUAI retaining project-level equity. Dilution is likely and I already bake it into my share count, but the real risk is not dilution by itself. The risk is financing on bad terms. The game here is simple: retain as much project-level equity as possible while minimizing shareholder dilution and structuring incentives so the tenant, capital providers, and builders are all pulling in the same direction.

Other Risks

• Tenant: If they do not sign an anchor, the thesis breaks and the stock likely reprices.

• Execution: Permitting, equipment lead times, and construction complexity can slip timelines and add cost.

• Financing terms: Expensive capital or unfavorable structure can cap equity upside even in a “success” outcome.

• AI demand slowdown: A meaningful slowdown could reduce urgency, pricing, or timing for new capacity.

• New Mexico lawsuit: The New Mexico AG lawsuit was filed on 12/29/2025 over legacy NM oil and gas wells and allegations tied to ownership and transfer activity plus plugging liabilities. Since then, two factors matter for how I frame materiality to TCDC. First, NUAI announced its partnership with PDI on 1/16/2026, after the lawsuit was already public. Second, management reiterated that an anchor hyperscale tenant is still expected before the end of Q1 2026. I do not dismiss the lawsuit as noise, because hyperscalers and infrastructure capital care about reputation and counterparty risk. But the sequencing is important: PDI chose to partner post-lawsuit, which suggests they did not view the NM overhang as fatal to leasing or financing the Texas project. Still, I give zero credit to the ~3,500-acre New Mexico site in my valuation.

Conclusion:

At a ~$400M market cap, NUAI is priced like the market is assigning a low probability to TCDC becoming a financed, contracted, gigawatt-scale campus. Based on my success/failure cases, the current price implies roughly a ~14% probability of success. My view is that the market is underestimating that probability.

This is a credibility trade. Since the New Mexico lawsuit became public, NUAI announced the partnership with PDI and management reiterated the target for an anchor hyperscaler announcement before the end of Q1 2026. Those are not guarantees, but they are signals that serious counterparties are still willing to engage. If NUAI secures a creditworthy anchor, the narrative shifts from micro-cap option to underwriting contracted infrastructure cash flows.

On my probability-weighted NOI / cap-rate model, that gets me to ~$16.50 per share (over 2x upside) with just a 60% implied success probability. The bet is not that everything goes right. It’s that the probability of success is meaningfully higher than ~14%, and the market re-prices accordingly.

There is also a real bull case if NUAI crosses the APLD-type threshold in the market’s eyes. My APLD MW framework implies ~$70 per share (nearly 10x upside) if NUAI is valued like a scaled, derisked, powered-land-and-shell developer. That outcome likely requires an anchor tenant, a credible financing stack, and early proof of execution. I am not underwriting perfection, but this gap highlights the explosive re-rating potential once the market stops viewing NUAI as a speculative concept and starts valuing it as an AI power platform.

Share

Leave a comment

Disclosure: I am long NUAI. I may buy, sell, or change my position at any time without notice. This article reflects my personal opinions and is for informational and educational purposes only. Nothing in this article should be construed as investment advice or a recommendation to buy or sell any security. I have not received compensation from NUAI, its management, or any affiliated party for this content. I have no business relationship with the company.