The received wisdom about SanDisk is that it got lucky on a commodity. NAND flash went short, spot prices ripped, and a spun-out memory business rode the cycle from a $29.62 low in April 2025 to $1,641.11 at the close on 14 August 2026. Tidy story. It is also the wrong one, and getting it wrong is why most investors will misread Fermi (NASDAQ: FRMI) at $6.40. What actually re-rated SanDisk was not the price of a bit. It was the conversion of spot exposure into contracted exposure — and on 13 August the company put numbers on exactly that, disclosing eight customers under long-term agreements covering roughly half of FY2027 bits. Fermi executed the first step of that identical conversion three days earlier, signing a 15-year, ~$6.5bn binding lease with TensorWave. The market has not repriced it. Eight analysts carry a $17.50 consensus target against a $4.10bn market cap.
The insight: multiples follow contracts, not commodities
Having tracked the memory complex through the whole of this cycle — the $8.97bn quarter that still sold off 8%, the Western Digital sibling trade, the peak-cycle fear that dominated the investor day preview — the single most instructive fact is this: SanDisk’s multiple did not expand when NAND prices rose. It expanded when the revenue stopped being spot.
At its 2026 Investor Day, SanDisk laid out an FY2028–FY2030 model of mid-to-high-teens revenue growth, roughly 80% non-GAAP gross margin, about 75% operating margin and around 50% adjusted free cash flow margin. Nobody underwrites those numbers off a commodity print. They are underwritten by the structure sitting beneath them: eight customers signed to what SanDisk calls New Business Model agreements, built on committed volumes, enforceable contractual frameworks and minimum financial guarantees, covering approximately 50% of bits in FY2027 and approximately two-thirds in FY2028. That is a cyclical business buying its way out of cyclicality. The re-rating is the market paying an annuity multiple for what it used to price as a commodity.
Now hold that template against Fermi. On 10 August the company signed its first binding lease at Project Matador: 222 MW of total facility power to TensorWave, a 15-year initial term with two five-year renewal options, approximately $6.5bn of contracted revenue, and expansion rights across two further data centres taking the site past 650 MW. Strip the sector labels and the two events are the same event. A capital-intensive asset with volatile, unbankable spot economics converts a slice of its output into long-duration contracted cash flow. SanDisk is roughly two-thirds of the way through that conversion and trades accordingly. Fermi has done it once.
Key facts
- FRMI last close $6.40, down 82.7% from its $36.99 intraday debut high — Fermi Inc, 14 August 2026 (StockAnalysis)
- Consensus price target $17.50 across 8 analysts; median $14.50, high $35.00, low $6.00 — StockAnalysis, August 2026
- ~$6.5bn contracted revenue over a 15-year initial term from the TensorWave lease, 222 MW phase 1 — Fermi press release, 10 August 2026
- ~6 GW of 17 GW planned capacity already permitted, with over $1.5bn invested in buildout to date — Fermi, 10 August 2026
- $92m cash at Q2 2026 against $520m net debt pre-offering; $431m convertible notes at 5.00% due July 2031 — Fermi Q2 2026
- SanDisk: 8 customers under long-term agreements covering ~50% of FY2027 bits and ~two-thirds of FY2028 bits — SanDisk Investor Day, 13 August 2026
- SNDK $1,641.11, from a $29.62 closing low in April 2025 — a 55x move in 16 months (StockAnalysis)
What is actually being built, and why the site matters more than the buildings
Project Matador sits on roughly 5,769 acres in Carson County, north-east of Amarillo, Texas. The design is behind-the-meter: rather than queue for a grid interconnection, Fermi intends to generate on campus, combining natural gas, nuclear, solar and battery storage into what it markets as a private “HyperGrid.” Construction is visible rather than theoretical — 4.6 miles of natural gas lines, 11.3 miles of perimeter fencing, 7.2 miles of water distribution, and roughly 300 acres of land prepared. Federal air permits are secured for 6 GW with a further 5 GW filed.
The gas turbine strategy is deliberately staged around delivery lead times, which is the part most write-ups skip. Seven GE TM2500 mobile units (126 MW) carry a two-month delivery window. Three GE Vernova FR6B units (116 MW) come in eight months. Six Siemens SGT-800s land at fifteen months, with the larger SGT6-5000F frames at twenty-two. That ladder is what makes a 210 MW target for 1 July 2027 and roughly 640 MW by Q4 2027 credible rather than aspirational: the early megawatts arrive on mobile units while the heavy frames are still in transit.
Capital structure is where it gets genuinely interesting. Under a build-own-operate-transfer arrangement with the Hillcore Alliance, Hillcore finances, constructs and operates 2.6 GW of combined-cycle generation with zero capital required from Fermi. Fermi becomes anchor offtaker under a 20-year power purchase agreement with 10-year renewal increments, and holds an option — not an obligation — to acquire the assets at fair market value after year ten. A company with $92m of cash does not build 2.6 GW. It rents it, and keeps the call option.
The counterparty was explicit about why it signed. “Power is the critical constraint in AI infrastructure, and the vision and scope of what Fermi is building at Project Matador resonates with our own,” said Darrick Horton, CEO and Co-Founder of TensorWave. Fermi chairman Marius Haas framed it more plainly: “A lease of this size and this term is a tremendous vote of confidence in Fermi.”
Who is actually doing what: the industry response
The useful test of any “power is the bottleneck” thesis is whether the people who buy compute are behaving as though it is true. They are. In August, Amazon confirmed plans for a large dedicated gas power plant to supply a new AI data centre — when the largest buyer of compute on earth starts commissioning its own turbines rather than waiting in an interconnection queue, the constraint has demonstrably moved off the wafer and onto the electron. That is simultaneously the strongest validation of Fermi’s thesis and its most serious long-run competitive threat: hyperscalers that self-supply do not need a merchant campus.
TensorWave itself deserves scrutiny rather than applause, because it is the load-bearing wall of the entire bull case. It is an AI cloud that offers AMD GPUs exclusively, and it raised $350m in June 2026 at a reported $1.55bn valuation, co-led by AMD Ventures and Magnetar, after a $100m Series A in May 2025. The lease will house tens of thousands of next-generation AMD Instinct GPUs. Read that carefully: a company valued at roughly $1.55bn has signed a 15-year obligation of approximately $6.5bn. The contracted revenue is only as good as the counterparty’s ability to pay it across a decade and a half, and no amount of contract length fixes a thin credit. This is the honest gap in the SanDisk parallel — SanDisk’s eight NBM customers are, overwhelmingly, established hyperscalers and OEMs. Fermi’s one customer is a venture-backed startup.
On the equipment side, the incumbents are quietly confirming the same demand picture. Power and thermal infrastructure names have been repricing all year on data centre capex — the Vertiv setup is the cleanest read-through in the listed space. Fermi’s turbine order book with GE Vernova and Siemens Energy is a small, verifiable piece of that same flow.
The numbers: what $6.5bn actually implies
Run the arithmetic the market has not yet run. Approximately $6.5bn spread across a 15-year initial term is roughly $433m of contracted annual revenue. Against a $4.10bn market capitalisation on 640.47m shares, that single lease represents about 10.6% of the entire equity value in annual contracted revenue — from 222 MW. Per unit, the lease prices at roughly $29.3m per MW across the term, or about $1.95m per MW per year.
The leverage sits in what remains uncontracted. That 222 MW is about 3.7% of the 6 GW already permitted, and roughly 1.3% of the 17 GW long-term ambition. Applying the same per-MW economics to the already-permitted 6 GW is an arithmetic illustration rather than a forecast — it assumes flawless execution, unlimited demand and stable pricing, none of which are safe — but it frames why the $35 street high exists at all. The gap between the $6.00 low target and the $35.00 high is not a disagreement about this year’s earnings. There are no earnings. It is a disagreement about how many of the remaining megawatts get contracted, and to whom.
The near-term arithmetic is considerably less romantic. Fermi lost $26m in Q2 2026, or $0.04 a share, and burned $49m in operations — though that burn was down 50% quarter-on-quarter. Trailing twelve-month net income is -$737.9m against EPS of -$1.25. Revenue does not begin until late Q3 or early Q4 2027. That is more than a year of pure cash consumption before the first contracted dollar arrives, funded by $431m of convertible notes at a 5.00% coupon maturing July 2031, struck at a ~$9.52 conversion price with a $34.5m capped call limiting dilution to roughly 2% even if the stock triples. Net proceeds were $417m, with no financial maintenance covenants. That $9.52 conversion price is worth remembering: it is the most honest near-term line in the sand anyone has drawn on this stock, and it sits 49% above spot.
Permission to build is the real scarce asset
The regulatory tension in this story is not chips or capital. It is consent. The most-discussed data centre item on Hacker News in the past month was research showing most Americans say “not in my backyard” to AI data centres — 146 points and 309 comments — followed by a Wall Street Journal piece on a rural community that turned down $26m rather than host one. Local opposition has become the binding permitting constraint across the sector, and it is almost entirely absent from the models.
This is where Fermi’s least glamorous asset does the most work. An already-assembled 5,769-acre tract in sparsely populated Carson County, with roughly 6 GW of federal air permits already secured and another 5 GW filed, is a categorically different asset from a proposed campus outside a suburb. Texas offers a permissive state posture, and Texas Tech University System chancellor Brandon Creighton publicly welcomed the lease, saying “TensorWave’s selection of Fermi sends a strong signal about the enormous potential of this project.” The behind-the-meter design also sidesteps the multi-year ERCOT interconnection queue that gates conventional grid-connected projects.
The nuclear component is the piece to discount hardest. Fermi markets Matador as hosting one of the largest new nuclear complexes in America, but nuclear licensing runs on regulatory timescales measured in years and no NRC milestone featured in the Q2 disclosures. Every megawatt that matters before 2030 is gas, solar or storage. Investors paying for the nuclear narrative today are paying for an option with a very long expiry.
Governance: the overhang that just lifted
Any honest FRMI thesis has to deal with the year the company has had. Co-founder Toby Neugebauer was removed as CEO and subsequently terminated for cause, then ran a proxy campaign to call a special meeting and install John Sellers as chairman and CEO. He suspended it after securing consents from approximately 31.0% of outstanding shares against revocations of approximately 36.4%, one day after ISS recommended shareholders withhold consent. The seat sat empty for more than three months before Lee McIntire — formerly of CH2M Hill and of Bill Gates’ nuclear venture TerraPower — became CEO effective 11 August 2026.
Governance chaos of that magnitude is precisely why a stock trades at a third of its consensus target. It is also, mechanically, why the setup is interesting now: the proxy fight is resolved, the CEO seat is filled by an infrastructure builder, and the first binding lease landed within a fortnight of both. SanDisk’s own April 2025 low came when the market had written it off as a commodity business its parent had discarded. Assets get cheap when the story is embarrassing.
What happens next
Prediction one: the second tenant is the catalyst, not the first. One lease can be dismissed as a favourable deal struck with a friendly, venture-backed counterparty. Two leases establish that Matador has a repeatable commercial motion and that the per-MW pricing is a market rate rather than an anchor negotiation. SanDisk did not re-rate on customer one; it re-rated as the NBM count marched toward eight. Expect the next announcement to move FRMI more than this one did.
Prediction two: first power slippage is the highest-probability disappointment. Fermi targets first power in 2026 and roughly 210 MW by 1 July 2027. The mobile-turbine ladder makes those dates defensible, but a company that has never energised a campus is forecasting one on a schedule with no slack. Any slip pushes the late-2027 revenue start into 2028 and forces another financing at a lower price — the single most plausible path to the $6.00 street low.
Prediction three: credit quality of the offtake book becomes the whole argument by 2027. If Fermi’s next counterparties are investment-grade hyperscalers rather than venture-backed clouds, the contracted revenue starts deserving an infrastructure multiple and the SanDisk analogy holds. If the book stays concentrated in thinly capitalised AI startups, the market will keep discounting the contracts no matter how long their stated terms — and it will be right to.
The honest summary is that Fermi at $6.40 is not a cheap version of SanDisk. It is SanDisk in April 2025: unloved, structurally loss-making, governance-scarred, and sitting on an asset whose scarcity the market has not yet agreed to pay for. The difference between the $6.00 bear target and the $35.00 bull target is roughly fourteen months of execution. That is precisely how long SanDisk’s own re-rating took to become obvious, and by the time it was obvious the stock had already done most of it. For the fuller memory-side picture, our SanDisk bull and bear case maps where that trade now stands.
Frequently asked questions
What is Fermi’s possible price target?
The consensus price target across eight covering analysts is $17.50, implying roughly 173% upside from the $6.40 close on 14 August 2026. The median is $14.50, the street high is $35.00 and the street low is $6.00. The consensus rating is Buy, though the spread between the high and low targets is unusually wide because the company has no revenue yet and valuation depends almost entirely on how much of its permitted capacity gets contracted.
Why is FRMI stock down more than 80% from its high?
Fermi listed on 1 October 2025 and traded as high as $36.99 intraday on debut before falling to $6.40. The decline reflects a post-IPO valuation reset, a long pre-revenue runway with revenue not starting until late 2027, heavy cash burn, and a damaging governance fight in which co-founder Toby Neugebauer was terminated for cause and then ran an unsuccessful proxy campaign. The CEO seat was vacant for over three months.
How is Fermi similar to SanDisk?
Both are capital-intensive businesses in structurally short markets that the market initially priced as speculative or commodity exposure. SanDisk re-rated once it converted spot volume into long-term contracted agreements with committed volumes and minimum guarantees. Fermi has just executed the first version of that same conversion, turning speculative megawatts into a 15-year, ~$6.5bn contracted lease. The mechanism is contract structure, not commodity price.
Who is TensorWave and can it pay a $6.5bn lease?
TensorWave is an AI cloud provider offering AMD GPUs exclusively, backed by AMD Ventures and Magnetar. It raised $350m in June 2026 at a reported $1.55bn valuation. The obvious concern is that a company valued near $1.55bn has committed to roughly $6.5bn of payments over 15 years. Counterparty credit quality is the single largest unquantified risk in the Fermi bull case and deserves more scrutiny than the headline number gets.
Is Project Matador actually a nuclear project?
Not in any timeframe that affects the next several years. Fermi markets Matador as integrating nuclear alongside natural gas, solar and battery storage, but no NRC licensing milestone appeared in the Q2 2026 disclosures, and nuclear licensing runs on multi-year regulatory timescales. Every megawatt scheduled before 2030 comes from gas turbines, solar and storage. The nuclear element is best treated as a long-dated option rather than near-term capacity.
What would make the bear case right?
Three things. A slip in first power or the July 2027 210 MW milestone, which would push revenue into 2028 and force a dilutive financing. A failure to sign a second major tenant, which would leave the campus dependent on one thinly capitalised counterparty. Or a broader shift where hyperscalers self-supply power — as Amazon’s own gas plant plans suggest — removing the need for merchant AI power campuses altogether. UBS already carries a $6.00 target on essentially this reasoning.
This article is for information only and is not investment advice. Prices and analyst targets are as of the close on 14 August 2026 and will have changed.
