SharkWater Trading  •  Digital Infrastructure • Power • Sector Map

Two Ways To Own The AI Buildout, And Only One Of Them Needs A Tenant

September 7, 2026

Bottom Line Up Front

Everything called "AI data center infrastructure" is really two different businesses stacked on top of each other. Layer one is the powered shell landlord, mostly converted bitcoin miners, whose entire value turns on whether a creditworthy tenant signs. TeraWulf has 401 MW to Anthropic on a 20 year lease worth roughly $19 billion. Core Scientific reports $24 billion or more of contracted capacity. Cipher has two hyperscaler leases. Keel has zero signed leases as of this writing.

Layer two is the supplier, and it gets paid on aggregate build volume rather than on any one landlord winning. GE Vernova closed Q2 2026 with a $176 billion backlog and $24.2 billion of quarterly orders, up 88 percent organically. Quanta carries a $53.4 billion backlog. The risk is not the same risk. Layer one is a binary on a signature. Layer two is a bet on volume, priced accordingly.

The Stack, In Plain Terms

The binding constraint on AI compute right now is not silicon. It is energized land with a live grid interconnect. That is why a cohort of bitcoin miners suddenly became infrastructure companies. They already owned the scarce thing.

Layer one takes that power position and rents it as a powered shell or a colocation contract. Layer two sells the turbines, switchgear, transformers, chillers, and field labor that turn a permitted site into a building full of racks. Layer two invoices whether or not any individual layer one developer ever finds a tenant.

A landlord with no lease is a finished dock with no boat tied to it. The pilings are driven, the water is deep, the slip is legally yours. It earns nothing until somebody throws a line.

Layer One: The Landlords

Rank these by signed contracts, not by pipeline gigawatts. Pipeline is a claim. A lease is a filing.

Name	Contracted Status	Counterparty
WULF TeraWulf	401 MW critical IT, 20 yr, ~$19B; up to ~$33B with both 5 yr extensions. FY2025 leases totaled 522 critical IT MW.	Anthropic
CORZ Core Scientific	~590 MW, 12 yr take or pay, $10B+ revenue potential, 80% to 85% anticipated margin. $24B+ total contracted capacity.	CoreWeave, others
CIFR Cipher Digital	300 MW gross to Fluidstack (207 MW critical IT), 10 yr, ~$3.8B, with $1.73B Google backstop. Separate 15 yr, 300 MW AWS lease.	Fluidstack / Google, AWS
IREN IREN Limited	Not a lease. A 5 yr GPU cloud services contract, ~$9.7B through 2031, 20% prepaid, 200 MW critical IT at Childress. IREN also bought ~$5.8B of GPUs from Dell.	Microsoft
KEEL Keel Infrastructure	No signed lease. Three priority sites near full permitting, $819M liquidity as of Aug 7, 2026, negative $23.7M adjusted EBITDA per quarter.	None yet

Sources: TeraWulf Q2 2026 Form 8-K and Form 10-Q (filed 2026); Core Scientific Q2 FY26 earnings presentation, EX-99.2 to Form 8-K dated July 28, 2026; Cipher Mining Form 8-K exhibits dated September 25, 2025 and November 20, 2025; IREN Form 8-K exhibit dated November 3, 2025; Keel Infrastructure Q2 2026 results release dated August 10, 2026. The dollar value of the Cipher AWS lease is not stated in the Cipher release I reviewed and is carried in secondary coverage only. See data gaps below.

Note the IREN row carefully, because aggregators file it next to the others and it does not belong there. TeraWulf, Core Scientific, and Cipher are landlords. IREN sells compute. It owns the data center and the GPUs, which is why its headline contract value is so much larger per megawatt, and also why it carries hardware refresh risk the landlords do not.

What A Megawatt Actually Rents For

Headline contract values are not comparable until you divide by capacity and term. Do that and the spread gets interesting.

Contract	Value	MW	Term	$M per MW-yr
WULF / Anthropic	$19.0B	401	20 yr	2.37
CIFR / Fluidstack	$3.8B	207	10 yr	1.84
CORZ / CoreWeave	$10.0B	590	12 yr	1.41
IREN / Microsoft, gross	$9.7B	200	5 yr	9.70
IREN / Microsoft, net of GPU capex	$3.9B	200	5 yr	3.90
KEEL	None	0	n/a	0.00

DERIVED BY SHARKWATER. Simple division of stated contract value by stated capacity and initial term. Not a company reported metric. No discounting, no ramp schedule, no escalators. Megawatt definitions are not uniform across issuers: WULF, CIFR, and IREN figures reference critical IT load, the CORZ figure references leased power across five sites. Cross-check: Core Scientific separately discloses roughly $850M average annual colocation GAAP revenue on the CoreWeave contract, which computes to $1.44M per MW-yr against $1.41M above. IREN separately discloses ~$1.94B targeted annualized run rate on 200 MW, which computes to $9.70M per MW-yr and matches exactly.

The takeaway is not that TeraWulf negotiated better than Core Scientific. It is that a twenty year lease to a frontier lab and a twelve year take or pay to a neocloud are different instruments with different credit behind them, and the per megawatt spread is roughly 68 percent. When someone tells you a developer has "2 gigawatts of pipeline," that number is worth nothing until you know what a megawatt rents for and who is signing.

Layer Two: The Suppliers

These companies do not need to know which landlord wins. They ship into the aggregate.

Function	Names	Hard Number
Power generation	GEV, CAT	GE Vernova Q2 2026: $176B total backlog, $24.2B orders up 88% organically, gas equipment backlog and slot reservations 100 GW to 116 GW in one quarter, targeting 125 GW by year end. Electrification data center orders above $5B year to date.
Electrical distribution	ETN, POWL, HUBB, NVT	Eaton Q2 2026 revenue $8.53B, up 21.4%, adjusted EPS $3.15.
Power and cooling in the white space	VRT, MOD, TT, AAON	Vertiv 2026 guidance as of July 29: net sales $13.8B to $14.2B, adjusted diluted EPS $6.65 to $6.75. Deferred revenue $1.815B at Dec 31, 2025 to $3.634B at Jun 30, 2026.
Field construction and MEP	PWR, EME, FIX, MTZ, DY	Quanta Q2 2026: revenue $9.56B, record backlog $53.4B, RPO $33.6B, 2026 revenue guidance raised to $39.3B to $39.7B.
Baseload power owners	CEG, VST, TLN	Signing PPAs directly with hyperscalers to bypass interconnect queues.

Sources: GE Vernova (NYSE: GEV) second quarter 2026 Form 8-K dated July 22, 2026; Quanta Services (NYSE: PWR) second quarter 2026 results, Exhibit 99.1 to Form 8-K dated July 30, 2026; Vertiv (NYSE: VRT) guidance update dated July 29, 2026 and June 30, 2026 balance sheet. Eaton (NYSE: ETN) Q2 2026 figures are from secondary coverage of the July 31, 2026 release and were not verified against the filing. Baseload power row is descriptive and carries no cited figure.

The Vertiv Line Nobody Read

Vertiv is the cleanest single expression of the buildout, at roughly three quarters of revenue from data center customers. It is also the best illustration of why headline numbers deserve suspicion.

In Q2 2025 the company led its release with backlog: $8.5 billion, book to bill about 1.2 times. Entering 2026 it was pointing at roughly $15 billion. The Q2 2026 release, published July 29, contains no backlog figure and no orders figure at all. Revenue of $3.274 billion came in under consensus, and the stock fell as much as 17 percent that day.

Here is the line three rows below the headline. Deferred revenue, which is cash customers have already paid for equipment not yet delivered, went from $1.815 billion at December 31, 2025 to $3.634 billion at June 30, 2026. Backlog is a management defined metric. Customer cash on the balance sheet is not. Those two facts point in opposite directions and the second one is audited.

I am not resolving that for you. I am telling you the disclosure changed and that the reappearance of a backlog figure in the Q3 release is the single most informative thing Vertiv can publish.

The Bull Case

• The contracts are long, large, and increasingly credit enhanced. Twenty year terms, take or pay structures, Google backstopping $1.73 billion of a tenant's obligations at Cipher. This is not spot GPU rental. It is contracted infrastructure cash flow.
• Supplier backlogs give multi year visibility that does not depend on picking a winner. A $176 billion GE Vernova backlog and a $53.4 billion Quanta backlog are orders already booked, not a forecast of orders.
• Power is a genuine physical bottleneck and it has a lead time. GE Vernova is booking gas turbine slots into 2031. Interconnect positions take years and litigation to replicate. Scarcity is not narrative here, it is queue position.
• Hyperscalers keep reaching outside their own footprints. Microsoft, AWS, Google, and Anthropic have all signed with third party developers rather than build everything first party. That is revealed preference about how tight capacity is.

The Bear Case

• Layer one without a lease is a cash burning developer. Keel runs negative $23.7 million adjusted EBITDA per quarter before development capital and has declined to restate its three lease target. The liquidity buys time to sign, not time to build.
• There is a duration mismatch buried in the stack. Landlords underwrite seven to nine year paybacks and sign ten to twenty year leases. The neoclouds sitting between them and the end demand run GPU contracts of two to five years. Somebody is carrying that gap.
• Supplier multiples already discount the visibility. Valuations across the electrical and cooling names sit near historical highs. You are paying for the backlog, which means the backlog has to convert on schedule.
• Disclosure quality is deteriorating at the edges. Vertiv stopped printing backlog. Developers publish pipeline gigawatts that mix energized capacity with load studies. When companies change what they show you, that is information.

The SharkWater Take

I would rather own the picks than the claims, and I would rather own a contracted landlord than a hopeful one. Those are two separate judgments and I hold both.

Layer two is where the buildout gets expressed without a counterparty bet. GE Vernova and Quanta have booked orders in hand and a physical lead time protecting the position. That does not make them cheap. It makes the failure mode a slower schedule rather than a zero, and slower schedule is a survivable outcome. That is the trade I would size normally.

Inside layer one I draw a hard line at the signature. TeraWulf, Core Scientific, and Cipher have executed documents with named counterparties, disclosed terms, and in Cipher's case a backstop from Google. Those are infrastructure businesses with an execution problem. Keel is a real estate option with a burn rate. It may well work, and I said so when I wrote it up in August, but it is a different instrument and it deserves a different position size. Do not let a shared sector label collapse that distinction.

And I would not treat IREN as a landlord comp. Its $9.7 billion headline is nine times the per megawatt economics of a Core Scientific lease because it includes the GPUs. Net of the $5.8 billion Dell purchase, the number is closer to $3.9 million per megawatt year, still the best in the group, earned by taking hardware obsolescence risk onto its own balance sheet. That may be a fine trade. It is not the same trade.

What I Could Not Reconcile

• Cipher AWS contract value. Secondary coverage cites roughly $5.5 billion over 15 years. The Cipher release I reviewed states 300 MW of capacity delivered in 2026 but not a dollar figure. I left it out of the derived table rather than compute against an unverified number.
• Megawatt definitions are not uniform. "300 MW" at Cipher's AWS site is not clearly stated as critical IT load versus gross capacity. At Barber Lake the same 300 MW gross corresponds to 207 MW of critical IT. That distinction moves per megawatt math by more than 40 percent and issuers are inconsistent about it.
• Eaton's data center backlog. A secondary source cites 307 GW of US data center backlog representing 15 years of work at current build rates. I could not verify that against the filing and have excluded it. The Q2 revenue and EPS figures above are also secondary.
• A Core Scientific AMD agreement. A July 28, 2026 8-K summary references a 2.5 GW AI capacity agreement with AMD. I did not read the underlying terms and have not included it in any figure here.
• Keel's lease status. "No signed lease" reflects the absence of any announcement I could locate through September 7, 2026. It is not a confirmed EDGAR negative.
• No prices, market caps, or valuation multiples appear in this post. Any I could source would be stale by the time you read it, and the argument does not need them.

Tight lines. — SharkWater

Educational and informational purposes only. Not personalized investment advice. All figures sourced as noted and accurate as of publication. Contracted revenue figures are management disclosures subject to delivery, commissioning, and counterparty performance, and are not guaranteed. Per megawatt calculations are the author's arithmetic on stated figures, not company reported metrics. The author may hold positions in securities discussed. Do your own work.

How To Weaponize Real Estate Against Your W-2 Tax Liability

SharkWater Trading  •  Strategy Desk • Wealth Preservation • Real Estate

The Phantom Expense: How To Weaponize Real Estate Against Your W-2 Tax Liability

September 6, 2026

Bottom Line Up Front

The One Big Beautiful Bill Act (P.L. 119-21, enacted July 4, 2025) killed the bonus depreciation phasedown and restored 100 percent first-year expensing permanently for qualified property acquired and placed in service after January 19, 2025. Under the old schedule this strategy would have been running at 20 percent bonus in 2026 and zero in 2027. It is running at full strength instead. On a modeled $1.2 million short-term rental with a cost segregation study reclassifying 25 percent of depreciable basis, Year One depreciation comes to $258,545 against roughly $12,000 of pre-depreciation cash flow, a $246,545 paper loss worth $91,222 at a 37 percent federal marginal rate.

Two things the promoters leave out. First, Section 461(l) caps how much of that loss reaches your salary, and the 2026 ceiling fell to $256,000 single and $512,000 joint from $313,000 and $626,000 in 2025. Second, cost segregation converts the deduction into Section 1245 property, which recaptures at ordinary rates up to 37 percent on exit rather than the 25 percent cap that applies to unrecaptured Section 1250 gain. The risk sits in the exit and in the hour log, not in the deduction.

Two Buckets, One Wall

Section 469 of the code sorts income into nonpassive and passive. Your salary is nonpassive. Rental real estate is passive by statutory default under Section 469(c)(2), which means losses from it sit on Form 8582 and wait for passive income to absorb them. They do not touch your W-2.

There is a narrow exception at Section 469(i) allowing $25,000 of rental loss against ordinary income, but it phases out between $100,000 and $150,000 of modified AGI. Anyone earning enough to care about this strategy is already past it. So the entire game is finding a legitimate way out of the passive bucket.

The Core Mechanism: Depreciation

The IRS lets you deduct the cost of the structure, never the land, over 27.5 years for residential rental property and 39 years for nonresidential. The deduction requires no cash outlay in the year you take it. That is what makes it a phantom expense: a property can produce positive cash flow and still print a taxable loss.

Straight-line at 27.5 years is slow. On $960,000 of depreciable basis that is $26,182 a year, and the mid-month convention cuts the first year down further depending on when the property goes into service. Useful, but it is not going to move a six-figure salary.

Accelerating It: Cost Segregation Plus Full Bonus

A cost segregation study brings in an engineer to decompose the building into components with shorter recovery periods. Appliances, carpet, cabinetry, specialty electrical, and land improvements such as paving and landscaping land in 5-, 7-, and 15-year buckets instead of the 27.5-year structure.

By itself that only reshuffles timing. The leverage comes from pairing it with Section 168(k) bonus depreciation, which lets you expense the entire reclassified amount in year one. This is where the law changed and where most published guidance is still wrong. Under TCJA the bonus percentage was stepping down 80, 60, 40, 20, and then to zero in 2027. OBBBA Section 70301 removed the placed-in-service deadline, repealed the phasedown, and set the rate at 100 percent with no sunset. Treasury and the IRS issued Notice 2026-11 on January 14, 2026 confirming the mechanics and directing taxpayers to the existing Reg. 1.168(k)-2 framework with the dates updated to January 19 and 20, 2025.

Check the date on anything you read about this. A meaningful share of live 2026 content still quotes 20 percent bonus depreciation for the current year. That figure is dead.

What The Arithmetic Actually Produces

Modeled acquisition: $1.2 million residential property placed in service in April, land allocated at 20 percent, cost segregation reclassifying 25 percent of depreciable basis into bonus-eligible components, $12,000 of net cash flow before depreciation.

Line	Without Cost Seg	With Cost Seg
Purchase price	$1,200,000	$1,200,000
Land (non-depreciable, 20%)	$240,000	$240,000
Depreciable basis	$960,000	$960,000
Reclassified to 5/7/15-yr	$0	$240,000
Bonus depreciation, Yr 1 @ 100%	$0	$240,000
Structure, 27.5-yr, April mid-month	$24,727	$18,545
Total Year 1 depreciation	$24,727	$258,545
Cash flow before depreciation	$12,000	$12,000
Year 1 paper loss	($12,727)	($246,545)
Federal benefit @ 37% marginal	$4,709	$91,222

MODELED BY SHARKWATER. Every figure in this table is an illustration, not a quote. The 20% land allocation, 25% reclassification ratio, $12,000 cash flow, and 37% marginal rate are assumptions chosen to make the mechanism legible. Real land allocations come from the county assessor or an appraisal. Real reclassification ratios come from an engineering study and vary widely by property type. Depreciation method per IRC 168 and 168(k) as amended by P.L. 119-21 Sec. 70301; IRS Notice 2026-11 (January 14, 2026).

Door One: Real Estate Professional Status

Section 469(c)(7) removes the per se passive treatment of rental real estate if you perform more than 750 hours of service in real property trades or businesses during the year, and more than half of all personal services you perform in any trade or business are in real property. You then still have to materially participate in each rental activity, or make the grouping election under Reg. 1.469-9(g) to treat them as one.

The detail that kills most joint filers: under Section 469(c)(7)(B) the tests must be met by one spouse individually. Hours do not pool. A couple where each spouse logs 500 hours qualifies for nothing. A couple where one spouse works 2,000 hours as an anesthesiologist and the other logs 900 hours on the portfolio does qualify, through the second spouse, because the first spouse's job is irrelevant to the second spouse's ratio.

This is one of the most litigated positions in the code. In Gragg the taxpayer cleared 750 hours but never separately proved material participation, and the losses stayed passive. In Sezonov v. Commissioner (T.C. Memo 2022-40) the work may well have been done, but there were no contemporaneous records. In Drocella v. Commissioner (T.C. Summ. Op. 2023-12) a couple with two full-time jobs and six rentals lost because they never documented their W-2 hours, so the denominator of the 50 percent test could not be verified.

Door Two: The Short-Term Rental Exception

If you cannot abandon a full-time career, there is a second route, and it does not run through Section 469(c)(7) at all. Temp. Reg. 1.469-1T(e)(3)(ii)(A) says an activity is not a rental activity if the average period of customer use is seven days or less. Fail to be a rental activity and you never enter the passive bucket by default in the first place.

That is only step one. The activity is then tested like any other business under the seven material participation tests at Temp. Reg. 1.469-5T. The two that STR owners actually use are the 500-hour test and the test requiring more than 100 hours with no other individual participating more than you. That second test is the one a property manager destroys. If a management company logs more hours than you do, you fail, and the study you paid for produces a suspended loss.

A trap worth flagging: because a qualifying STR is not a rental activity, a number of practitioners take the position that STR hours do not count toward the 750-hour REPS test either. I have not found a controlling authority settling this, and I am not going to present it as settled. Treat it as unresolved and get a written position from your CPA before relying on hours in both directions.

The Ceiling Nobody Puts In The Headline

Section 461(l) limits how much net business loss a noncorporate taxpayer can use against nonbusiness income in a single year. Everything above the threshold is disallowed, reported on Form 461, and converted into an NOL carryforward that is then subject to the 80 percent of taxable income limitation under Section 172.

OBBBA made this limitation permanent and reset the base year for indexing, which means the 2026 threshold went down, not up.

Tax Year	Single	Married Filing Jointly
2024	$305,000	$610,000
2025	$313,000	$626,000
2026	$256,000	$512,000

Source: Rev. Proc. 2023-34 (2024), Rev. Proc. 2024-40 (2025), Rev. Proc. 2025-32 Sec. 4.31 (2026), reflecting the P.L. 119-21 reset of the indexing base to 2024. A joint filer who could absorb $626,000 of net business loss in 2025 can absorb $512,000 in 2026 on identical facts.

The $246,545 loss in the model clears easily. Scale the strategy to a $3 million property or stack two acquisitions in the same year and you hit the wall, at which point the excess is deferred rather than lost, but the cash-flow math you underwrote no longer holds.

The Exit: What A 1031 Does And Does Not Bury

Section 1031 defers gain on the exchange of real property held for productive use in a trade or business or for investment. Since TCJA it applies to real property only. Personal property is out.

Two clocks, and they start on the same day. You identify replacement property in writing within 45 days of transferring the relinquished property, and you must receive the replacement by the earlier of 180 days or the due date of your return including extensions. That second half of the sentence catches people who close in November. The 45 days are inside the 180, not added to it.

A 1031 is bailing water from the bow compartment into the stern. The boat rides differently. It does not weigh less.

Here is the part that gets skipped. Cost segregation converts a chunk of your building into Section 1245 property, and Section 1245 depreciation recaptures as ordinary income at rates up to 37 percent, with no preferential cap. Unrecaptured Section 1250 gain on the structure is capped at 25 percent. So the acceleration you bought in Year One raises your recapture rate later. Section 1245(b)(4) limits the recapture recognized in a like-kind exchange to gain otherwise recognized plus the fair market value of replacement property received that is not Section 1245 property, which means the exposure carries into the replacement property rather than disappearing. If your replacement holds less Section 1245 property than what you gave up, some of it comes due at closing.

The stepped-up basis at death under Section 1014 is the only mechanism in this chain that eliminates rather than defers. Everything upstream of it is a loan from the Treasury.

Pricing The Deferral Honestly

Run the model as a financing transaction, which is what it is. You take $91,222 of federal benefit in Year One. On a ten-year hold and a taxable exit, the $240,000 of Section 1245 property recaptures at 37 percent for $88,800.

Discount Rate	PV of Yr-10 Recapture	Net PV Benefit
6%	$49,585	$39,215
8%	$41,132	$47,668
10%	$34,236	$54,564

DERIVED BY SHARKWATER from the modeled figures above. Assumes a taxable exit in year 10 with no 1031, a constant 37% marginal rate in both the deduction year and the recapture year, and ignores state income tax, the Section 1245 versus 1250 split within the reclassified components, the cost of the engineering study, and any change in law over the holding period. Not a projection.

Roughly 40 to 55 cents on the headline dollar, before the study fee and before state tax. That is a real number and it is worth having. It is not the number in the seminar.

On the study fee: quoted engagement costs vary enough by property size and complexity that I am not going to publish a range I cannot source to your specific property. Get a written quote before you underwrite the benefit, because the fee is a Year One cash cost against a Year One tax benefit.

The Bull Case

• The statutory window is permanently open. OBBBA Sec. 70301 removed the sunset entirely. Every prior version of this strategy carried a deadline that forced rushed acquisitions. That pressure is gone, which means you can underwrite the property on the property rather than on the calendar.
• The STR door does not require quitting your job. REPS is functionally closed to a full-time professional. Temp. Reg. 1.469-1T(e)(3)(ii)(A) plus the 100-hour material participation test is clearable by a motivated owner with one property and no management company.
• The regulation has survived two major rewrites. The seven-day exception is a 1988 temporary regulation that the IRS never finalized and never replaced. It came through TCJA in 2017 and OBBBA in 2025 untouched, and the Service litigates the facts rather than the rule.
• The benefit is front-loaded and the cost is back-loaded. Even at a punitive discount rate the deferral clears $39,000 of present value on the modeled property, and a disciplined operator can extend the hold or exchange indefinitely.

The Bear Case

• The deduction is not the hard part. The hour log is. Gragg, Sezonov, and Drocella all failed on documentation and denominators, not on the underlying concept. Reconstructed calendars produced after an audit notice have a poor record in Tax Court. The burden is on you.
• Section 461(l) got tighter in 2026, not looser. The joint threshold dropped $114,000 year over year. Anyone who modeled a large acquisition on 2025 figures is over-deducted on paper and holding a carryforward instead of a refund.
• Cost segregation raises your exit tax rate. You are trading a 25 percent capped unrecaptured 1250 rate for ordinary 1245 recapture at up to 37 percent. If your marginal rate is the same in both years, you have bought time value and nothing else.
• The service level that makes an STR profitable can make it a Schedule C business. Provide substantial services for guest convenience and the activity can leave Schedule E, at which point net profit picks up 15.3 percent self-employment tax. The passive-versus-nonpassive question and the self-employment question are decided separately, and winning the first does not settle the second.

The SharkWater Take

This works, and I think the STR door specifically is the most underused legal structure available to a high-earning W-2 professional right now. The permanence of 100 percent bonus is a genuine change and it removes the thing I disliked most about this trade, which was the artificial deadline pushing people into bad properties.

But I want to be precise about what you are buying, because the marketing around this is aggressive and the arithmetic is usually presented in one direction. You are not erasing tax. You are borrowing from the Treasury at zero stated interest against collateral you will surrender at ordinary rates. On the modeled property that loan is worth somewhere near $40,000 to $55,000 in present value on a ten-year hold. Worth doing. Not the number on the webinar slide.

The failure mode is not the IRS disallowing depreciation. It is buying a mediocre property because the tax math looked good, then discovering in month fourteen that a management company logged more hours than you did. Underwrite the asset first. If the property does not work as a property, the deduction is a subsidy on a bad trade.

Execution Notes

Sequencing: Placed-in-service date drives the mid-month convention and therefore Year One structure depreciation. Acquisition date must be after January 19, 2025 for the 100 percent rate under Notice 2026-11. Property acquired January 1 through 19, 2025 sits at 40 percent.

Documentation basis: Contemporaneous, per Reg. 1.469-5T(f). Date, property, specific task, hours, and a corroborating artifact such as an invoice, permit, or guest message. Start the log on day one, not in April of the following year.

Figures basis: All dollar amounts in this post are MODELED, not quoted. Statutory thresholds are sourced to Rev. Proc. 2025-32 and P.L. 119-21. Depreciation mechanics per IRS Notice 2026-11.

Unresolved: Whether hours spent on a qualifying STR count toward the 750-hour REPS test. Practitioner opinion is split and I found no controlling authority.

Invalidation: A property manager out-hours you on the 100-hour test. Average stay creeps above seven days. The 461(l) ceiling binds because you stacked acquisitions in one year. Your state decouples from federal bonus depreciation, which several do, including New York, New Jersey, Pennsylvania, and Illinois.

Tight lines. — SharkWater

Educational and informational purposes only. Not personalized investment, tax, or legal advice, and not a substitute for a CPA or attorney reviewing your specific facts. All dollar figures in this post are modeled illustrations, not quotes. Statutory and regulatory citations are accurate as of publication and tax law changes. The author may hold positions in securities or real property discussed. Do your own work.

DXZY Update

SharkWater Trading  •  Venture Desk • Closed-End Funds • NAV Mechanics

Destiny Tech100's NAV Rose 39.7 Percent Last Quarter. Its Largest Holding Is a Treasury Money Market Fund.

August 29, 2026

Bottom Line Up Front

Destiny Tech100 (NYSE: DXYZ) reported net asset value of $34.30 per share as of June 30, 2026, up from $24.56 at March 31. That is a 39.7 percent quarterly gain, and it will get read as portfolio performance. It is mostly not. The fund sold 17,191,674 new shares through its Jefferies at-the-market program during the quarter, roughly a 57 percent increase in the share count, at prices well above the prior NAV mark. Selling stock above NAV lifts NAV per share for everyone already holding it, and by my arithmetic that mechanism accounts for somewhere between 36 and 63 percent of the entire gain.

The range is that wide because the filing contradicts itself. It reports 17,191,674 shares sold at a weighted average price of $34.25, which is $588.8 million gross, and then reports net proceeds of $715,442,732. Net proceeds cannot exceed gross proceeds. One of those three numbers is wrong and the document does not say which.

Meanwhile the money went into Treasuries. 57.3 percent of the June 30 portfolio sits in the First American Treasury Obligations money market fund. The private venture book, the thing anyone buys DXYZ to own, is roughly $700 million of a $1.64 billion portfolio. I am not recommending a position here in either direction, and the last section explains why.

What Was Actually Filed

Destiny put out a Business Wire release at 8:11pm Eastern on Friday, August 28, announcing NAV of $34.30 per share for the quarter ended June 30, 2026, against approximately $1.64 billion of portfolio fair value. The same evening the fund filed Supplement No. 1 to its May 26, 2026 prospectus on Form 424B3, File No. 333-296216, which carries the same NAV figure plus the full position table.

The press release is four paragraphs. The filing is where the story is, and the two most important things in it are not in the release at all.

The First Thing: More Than Half the Fund Is Cash

The 424B3 position table lists First American Treasury Obligations, Class X, at 57.3 percent of portfolio. That is the single largest line by a wide margin. The next largest is Magnitude ANC III, an SPV with economic exposure to Anthropic Series B preferred, at 14.4 percent. Three separate SPVs carrying SpaceX Class A common stock add to 10.5 percent. Two OpenAI vehicles add to 2.6 percent.

So the fund's four biggest exposures, in order, are: Treasury bills, Anthropic, SpaceX, and OpenAI. Only three of those four are venture capital.

Compare the same table one quarter earlier. At March 31, 2026, on a portfolio of approximately $742.5 million, the money market line was 31.4 percent. Cash weight nearly doubled as a share of the fund in three months, and it did so because the at-the-market program was pumping capital in faster than the adviser was deploying it.

A boat rides higher when you pump the bilge dry. It does not ride higher because the water got deeper.

The Second Thing: Most of the NAV Gain Is Issuance

When a closed-end fund trades above NAV and issues new shares at that market price, every dollar raised above book value accrues to existing holders. It is a real economic benefit and it is entirely legitimate. It is also not investment performance, and it does not repeat once the premium closes.

DXYZ issued 17,191,674 shares in the quarter. The filing does not report shares outstanding or total net assets, so I derived the share counts below from the reported NAV per share and the reported approximate portfolio value. That derivation assumes net assets are close to portfolio value, which requires the fund to carry no material leverage. Treat every share count in this table as mine, not the fund's.

Line	Mar 31, 2026	Jun 30, 2026
NAV per share (reported)	$24.56	$34.30
Approx. portfolio value (reported)	$742.5M	$1.64B
Money market weight (reported)	31.4%	57.3%
Implied private book (derived)	~$509M	~$700M
Shares outstanding (derived)	~30.2M	~47.8M
ATM shares sold in quarter (reported)	8,489,359	17,191,674

Source: Form 424B3 Supplement No. 1 dated August 28, 2026 (File No. 333-296216) and Form 424B3 Supplement No. 6 dated May 12, 2026. NAV, portfolio value, money market weight, and ATM share counts are reported figures. Shares outstanding and private book value are DERIVED BY SHARKWATER from those reported figures and are not stated anywhere in the filings. The derivation assumes net assets approximate portfolio value.

Now the decomposition, and here is where the filing fights itself. Reported ATM detail for the quarter reads 17,191,674 shares at a weighted average price of $34.25, with net proceeds of $715,442,732. Multiply the shares by the price and you get $588.8 million gross. Net proceeds are reported $126.6 million higher than gross proceeds, which is arithmetically impossible.

I cannot tell you which figure is broken, so here is what the NAV bridge looks like under each.

Scenario	Capital raised	From issuance	From portfolio marks
Price times share count ($588.8M)	$588.8M	$3.51 / 36%	$6.23 / 64%
Stated net proceeds ($715.4M)	$715.4M	$6.18 / 63%	$3.56 / 37%

DERIVED BY SHARKWATER, not reported. Splits the $9.74 per share NAV increase between accretion from above-NAV share issuance and change in portfolio marks, under each of the two irreconcilable ATM figures in the August 28, 2026 424B3. Both scenarios rest on the derived share counts in the prior table.

The second scenario reconciles better with the rest of the document. A $168.8 million contribution from marks is close to the roughly $191 million by which the derived private book grew during the quarter, and that $191 million also includes whatever new private capital was deployed inside the quarter. The first scenario requires about $309 million of appreciation inside a private book that only grew $191 million in total, which does not fit unless there were substantial private sales the filing does not disclose.

My read is that the $715.4 million net proceeds figure is the reliable one and that the $34.25 weighted average price is understated, implying an effective issuance price near $41.62. But that is inference, not disclosure, and I am not going to present it as fact. What survives either way: a large minority to a clear majority of the headline NAV gain came from selling stock, not from owning it.

The Control Case: What This Looks Like Without an ATM

Fundrise Innovation Fund (NYSE: VCX) filed its Form NPORT-P for the same quarter end, and it is the cleanest possible comparison because VCX did not issue a single share.

Total net assets of $776,968 thousand at June 30, 2026, against $678,918 thousand at March 31. Divide each by the 35,797,138 shares outstanding reported in the March 31 audited N-CSR and you get exactly $21.70 and exactly $18.97, the two NAV figures the fund reports. The share count did not move. Every cent of that 14.4 percent NAV gain came from marks, and the filing itself shows the source: an $85,947 thousand net change in unrealized appreciation on Level 3 positions, plus $15,442 thousand realized.

Metric, June 30, 2026	DXYZ	VCX
NAV per share	$34.30	$21.70
Quarterly NAV change	+39.7%	+14.4%
Shares issued in quarter	17,191,674	0
NAV gain attributable to marks	37% to 64% (derived)	100%
Cash / money market weight	57.3% of portfolio	5.4% of net assets
Largest single exposure	Treasury MMF, 57.3%	Anthropic, over 20%
Leverage disclosed	None found in 424B3	$30.4M Barclays reverse repo

Sources: DXYZ Form 424B3 Supplement No. 1, August 28, 2026. VCX Form NPORT-P for the period ended June 30, 2026, and Form N-CSR for the period ended March 31, 2026. DENOMINATOR WARNING: the two funds do not use the same base. DXYZ states position weights as a percentage of approximate portfolio value. VCX states them as a percentage of net assets. On DXYZ's derived net assets the money market weight is roughly 57.8 percent, so the two are close here, but do not assume that holds in other quarters.

VCX is not the better fund because of this. It carries its own problems, including $30,400 thousand of Barclays reverse repurchase borrowing collateralized by data center CMBS, restricted securities at 82.2 percent of net assets, and single-name Anthropic exposure above 20 percent. The point is narrower. When you see a NAV number from a listed venture fund, the first question is whether the share count moved, and the answer changes what the number means.

What Happened to the Premium

For most of this year the retail conversation about DXYZ has been about its premium to NAV. Against the stale $24.56 mark, the fund looked like it was trading roughly 40 percent rich in late August. Against $34.30 it is trading close to flat.

The premium did not compress. The denominator moved. Nothing about the price had to change for a 40 percent premium to become no premium at all, and anyone who was short the premium and long the thesis that it had to close got the outcome without the mechanism.

This is the recurring lesson with quarterly-marked vehicles and I will keep repeating it. A premium computed against a mark that is two to five months old is a statement about the calendar, not about valuation. It tells you when the fund last opened its books. It does not tell you what the fund is worth today.

The Bull Case

• The accretion is real money. Issuing above NAV genuinely transfers value to existing holders. Whether it is $3.51 or $6.18 per share, it is not an accounting illusion, and holders through the quarter captured it.
• The cash is being deployed, and quickly. The filing discloses $169.0 million closed in three transactions after quarter end as of August 27: $150.0 million more into OpenAI on August 13, $15.0 million into Fluidstack on July 16, and $4.0 million into Boom Technology via a SAFE on August 4. That is real velocity against the September quarter.
• The top of the book is the right book. Anthropic at 14.4 percent, SpaceX at 10.5 percent across three vehicles, OpenAI at 2.6 percent before the August addition. If you want concentrated exposure to those three names in a liquid wrapper, the options are few.
• Buying near NAV is a different trade than buying at a 40 percent premium. Whatever the mechanism, the entry math available now is not the entry math that was available in June.

The Bear Case

• You are paying venture fees on a Treasury money market fund. That is the plain reading of a 57.3 percent cash weight. Every dollar in First American Treasury Obligations is a dollar you could hold yourself for a few basis points.
• The accretion engine only runs while the premium exists. If DXYZ is now near NAV, the largest single driver of last quarter's NAV growth is switched off. Nobody modeling forward growth off a 39.7 percent quarter should expect that rate to repeat.
• The SpaceX marks may reflect a price the fund cannot yet realize. SpaceX listed on June 12, 2026, inside the quarter. DXYZ holds it through three layered SPVs, one of which the filing describes as investing through multiple underlying SPVs with more than one layer. Public-price marks on locked-up, multiply-nested positions are the most fragile marks in the book.
• The filing does not reconcile. A prospectus supplement that reports net proceeds exceeding gross proceeds by $126.6 million is a disclosure quality problem, not a rounding problem, and it sits in the one paragraph a buyer most needs to trust.

The SharkWater Take

I am not taking a position and I am not telling you to take one, and the reason is specific rather than squeamish. The single most important input to any DXYZ trade right now is the premium or discount to NAV, and I cannot compute it to a standard I would put my own money behind. I have a hard, primary-sourced NAV of $34.30. I do not have a confirmed Friday closing price, and I do not have a reported share count, which means I do not have net assets either. Two of the three legs are derived and one of the derivations rests on a filing that contradicts itself.

What I will say plainly is that the headline is misleading, and predictably so. Over the next week you will see the 39.7 percent figure quoted as a venture return. It is not one. It is a fund that roughly doubled in size by selling stock at a premium, parked the proceeds in Treasuries, and reported the resulting per-share arithmetic as a quarterly result. The Wolfspeed post in July was the same shape in a different costume: a headline number that was an artifact of a corporate action rather than a description of the business.

The honest version of the DXYZ story is more interesting than the misleading one. A venture fund that is 57 percent cash, deploying $169 million in eight weeks into OpenAI and AI infrastructure, and that has just lost the premium that was funding its own NAV growth, is at a genuine inflection. I would rather write that post in three weeks with a confirmed share count than write a trade today off numbers I had to build myself.

Two things would change my mind and make this actionable. A reported shares outstanding figure and total net assets, which the semi-annual report should carry. And a corrected or clarified ATM disclosure. Until both land, this is a monitoring position, not a trade.

Data Gaps

• Shares outstanding and total net assets: NOT REPORTED. Neither figure appears in the August 28 424B3 or the press release. Every share count and net asset figure in this post is derived by me from NAV per share and approximate portfolio value, and assumes no material leverage.
• ATM proceeds: IRRECONCILABLE. 17,191,674 shares at a weighted average $34.25 is $588.8 million gross, against reported net proceeds of $715,442,732. Reported as filed. Not resolved.
• DXYZ closing price: NOT VERIFIED. No confirmed August 28, 2026 close was obtained from a primary or exchange source. No premium or discount figure in this post is computed against a live price, and the "close to flat" characterization is directional only.
• DXYZ leverage: NOT CONFIRMED EITHER WAY. The fund filed a 424B5 prospectus supplement for debt securities on May 26, 2026. Whether any debt was drawn as of June 30 is unknown, and if it was, the derived share counts here are overstated.
• VCX NPORT-P filing date: DISPUTED. One third-party account places it on EDGAR August 26, 2026, another August 28. The document content is confirmed; the accession timestamp is not.
• Private book values are derived, by subtracting the reported money market weight from the reported approximate portfolio value. The funds do not report a private-only subtotal.
• No options data appears in this post. No chain was pulled, no premium was modeled, and none should be inferred.

For educational and informational purposes only. Not personalized investment advice, and not a recommendation to buy or sell any security. Figures labeled DERIVED are SharkWater calculations, not issuer disclosures, and should be independently verified before use. Closed-end funds investing in privately held companies carry valuation, liquidity, and concentration risks that differ materially from those of listed equities. The author may hold positions in securities discussed. Do your own work.

 SharkWater Trading  •  Nuclear Desk • FERC Complaint • Interconnection Risk

Oklo's FERC Complaint Is Real. The Ohio Story Attached to It Might Not Be.

September 2, 2026

Bottom Line Up Front

Oklo Inc. (NYSE: OKLO) fell 5.03 percent to $38.53 on September 1, 2026, a day after PJM Interconnection's withdrawal of a 750 megawatt Oklo project from its study queue prompted an emergency complaint at FERC, docket EL26-101, filed August 28 and asking for a ruling by September 4. Nearly every outlet covering the story has tied that 750 megawatt project to Oklo's Meta-backed Pike County, Ohio campus. The complaint itself never says that.

The primary document names a Dominion-owned substation in Virginia, not a site in Ohio. That gap between what Oklo actually filed and the story built on top of it is where the real risk sits for anyone trading this on the headline alone.

What Oklo Actually Filed

The filing itself is public. Oklo Inc. v. PJM Interconnection, L.L.C., FERC Docket EL26-101, filed August 28, 2026, concerns a 750 megawatt project carrying queue number C01-1735, split across three fuel types: 150 megawatts of advanced nuclear, 300 megawatts of fuel cells, and 300 megawatts of natural gas. PJM pulled the project from its Cycle 01 interconnection study on August 3, 2026, citing Oklo's failure to demonstrate low-voltage ride-through stability across all three fuel types. Oklo's complaint asks FERC to reinstate the project at its original queue position and original Cycle 01 schedule, targeting a final interconnection agreement by May 19, 2028, rather than being pushed to Cycle 02 and a target of July 13, 2029. Oklo calls that roughly a 14-month delay and wants FERC to act on it by September 4, 2026.

The Ohio Story Nobody Can Source to the Filing

Coverage from Utility Dive, Yahoo Finance, and several others has assumed the withdrawn project is the 1.2 gigawatt Meta-backed campus in Pike County, Ohio that Oklo announced in January. The 750 megawatt figure lines up with that project's stated fuel mix, so the assumption is understandable. But the complaint text itself does not mention Ohio, Pike County, or Meta anywhere. The only site reference in the document is to a Possum Point 500 kilovolt substation bus, a real Dominion Energy asset located in Prince William County, Virginia, which is Dominion transmission territory, not the AEP or Duke footprint that covers southern Ohio.

That could mean the press correctly connected dots that exist in correspondence outside the public complaint. It could also mean this is a second, separate Oklo interconnection project, or that the Ohio link is simply wrong. As of this writing, no primary Oklo statement resolves which. No 8-K referencing the PJM dispute has been filed with the SEC.

Filing a FERC complaint over a lost queue position is like losing your spot in the harbor launch line. You can go argue with the harbor master, and you might even get it back. But arguing doesn't return the tide you missed while you were making your case.

The Balance Sheet Behind the Fight

Oklo's 10-Q for the quarter ended June 30, 2026, filed August 7, 2026, shows cash and equivalents of $1.645 billion plus marketable securities of $1.362 billion, a combined $3.01 billion. Net loss for the first half of 2026 was $81.6 million, with $65.5 million in cash used by operations over the same six months, roughly $32.8 million a month. Management's own language states the cash on hand is sufficient to fund operations for one year from the filing date. Revenue for the first half of 2026 was $1.21 million, described by the company as its first quarterly revenue period, largely from engineering and manufacturing services tied to recent acquisitions. This remains, by any conventional measure, a pre-revenue developer with a large cash cushion and a real burn rate, not yet a power company collecting checks.

The Contracts That Matter More Than One Queue Position

Two agreements anchor the growth story regardless of how the FERC complaint resolves. The Meta agreement, announced January 9, 2026, funds development and fuel procurement for the 1.2 gigawatt Pike County campus through prepayment, structured explicitly as advance funding rather than a traditional long-term power purchase agreement, with phase one targeted as early as 2030 and full buildout by 2034. The Switch agreement, from December 2024, is a non-binding master power agreement for up to 12 gigawatts through 2044, with individual binding contracts to follow as milestones are met. That 12 gigawatt figure shows up constantly in bullish commentary on the stock. It is not contracted revenue. It is a framework for future contracts that may or may not materialize on schedule.

Key Figures, Dated and Sourced

Figure	Value	Source
Close, Sept 1, 2026	$38.53 (-5.03%)	Market data, aggregator-sourced
Cash + marketable securities	$3.01B, as of 6/30/26	Form 10-Q, filed 8/7/26
H1 2026 net loss	$81.6M	Form 10-Q, filed 8/7/26
H1 2026 revenue	$1.21M	Form 10-Q, filed 8/7/26
Stated cash runway	1 year from 8/7/26	Form 10-Q, filed 8/7/26
FERC docket	EL26-101, filed 8/28/26	PJM-hosted filing PDF, FERC eLibrary
Withdrawn project capacity	750 MW (150 nuclear / 300 fuel cell / 300 gas)	FERC complaint, docket EL26-101
Market cap	~$7.19B, intraday 9/2/26	Aggregator (stockanalysis.com)
52-week range	$36.61 – $193.84	Aggregator (stockanalysis.com)

Market cap, close price, and 52-week range are aggregator-sourced (stockanalysis.com) and not exchange-primary. All filing-derived figures are sourced directly to the named SEC or FERC document. The Pike County, Ohio and Meta association with docket EL26-101 is press-reported, not confirmed in the primary complaint text, and should be treated as unverified.

The Bull Case

• The cash cushion is real. $3.01 billion against a burn rate near $33 million a month buys years of runway even without new financing, which is unusual for a company this early in commercialization.
• Licensing is moving faster than the industry norm. The NRC approved Oklo's principal design criteria topical report on May 6, 2026, in less than half the traditional review timeline, which should compress future licensing steps for the Aurora powerhouse design.
• The Meta funding isn't contingent on this one queue slot. The prepayment structure with Meta is separate from any single PJM interconnection filing, so a resolution against Oklo at FERC does not by itself unwind the underlying commercial relationship.

The Bear Case

• The headline narrative outruns the primary document. Trading this news as an Ohio, Meta-specific story is trading on an inference the filing itself does not support. If the connection turns out to be wrong, the thesis attached to today's move goes with it.
• The decline predates the FERC news. Oklo was already down roughly 43 percent year to date through August 31, 2026, and had posted new 52-week lows in the weeks before this complaint. The market was repricing this stock before PJM ever entered the picture.
• Revenue is not yet a company, it's a rounding error. $1.21 million over six months against a nine-figure quarterly burn means every headline gigawatt figure, contracted or not, is still years from showing up as cash in the door. First commercial power at Aurora-Idaho is not targeted before 2028.

The SharkWater Take

The FERC fight is real, dated, and worth watching into September 4. I am not trading Oklo on the Ohio story until somebody puts a primary document in front of me that actually says Pike County or Meta, because right now that connection is press inference sitting on top of a filing that names a Virginia substation. Strip the unverified narrative away and what's left is a well-funded, pre-revenue nuclear developer whose stock has already been cut in half this year on its own timeline risk, not on this complaint. That's a name to research on its execution against 2028 and 2030 milestones, not one to chase on a FERC deadline three days out.

Tight lines. — SharkWater

Educational and informational purposes only. Not personalized investment advice. All figures sourced as noted and accurate as of publication. Options involve substantial risk of loss. The author may hold positions in securities discussed. Do your own work.

 SharkWater Trading  •  Semiconductor Desk • Sell-Off • Customer Concentration

Cerebras Fell Three Times Faster Than the Chip Sector. No Filing Explains Why.

September 2, 2026

Cerebras Systems (Nasdaq: CBRS) closed September 1, 2026 at $172.62, down 6.31 percent, on a day the broader semiconductor tape fell 1 to 3 percent. Intel dropped 3 percent, Nvidia and AMD each fell 2 percent, Broadcom fell 1 percent, and the SOXX semiconductor index fell 2 percent, all attributed to a bond market selloff pushing the 10-year Treasury yield to 4.8 percent. No 8-K, no press release, and no filing of any kind explains Cerebras's extra move, and the stock was already bouncing roughly 6 percent intraday the next morning.

The thing actually worth tracking here isn't Tuesday's air pocket. It's a $25.4 billion contracted backlog whose customer concentration is migrating from two Abu Dhabi entities into one American one: OpenAI.

A Move the Sector Doesn't Explain

Every major chip name was down September 1, 2026, so a Cerebras decline on its own isn't a story. The size of it is. Against Intel's 3 percent, Nvidia and AMD's 2 percent each, Broadcom's 1 percent, and the SOXX index's 2 percent, Cerebras's 6.31 percent drop is two to three times the sector's move on a day with a clear, sector-wide cause already identified in the financial press. That gap is what makes this worth a post rather than a footnote.

A search of Cerebras's SEC filing history turns up nothing in the week before the drop beyond routine Form 4 insider-ownership updates on August 25 and 26. No 8-K. No S-1 amendment. Nothing on the company's own newsroom that surfaced in this research pass. By the next session, the stock had recovered a meaningful chunk of the move, trading back toward $183 intraday on September 2. A sharp, unexplained drop followed by a partial same-week reversal is the signature of a liquidity event, not new information.

A six percent drop on a two percent sector day isn't a storm front. It's a rogue wave. It moves through fast, and most of the time it doesn't tell you a thing about tomorrow's weather.

What's Actually New: Not Much, on the Record

The filings that do exist around this period are a cluster of Form 144 notices, insider intent-to-sell filings, dated August 17 through 21, 2026, alongside routine Form 4s and a Schedule 13D/A on August 18. None of these are unusual on their own. They line up with the first tranche of the company's IPO lockup, roughly 60 million shares, releasing around the August 12, 2026 earnings date. That's routine paperwork tied to a known calendar event, not a fresh catalyst for a Tuesday selloff.

The Real Story: A Loss That Isn't Really a Loss

Cerebras priced its IPO May 13, 2026 at $185 a share and began trading the next day on Nasdaq. Its second quarter, ended June 30, 2026, reported August 12, 2026, showed GAAP revenue of $180.1 million, up 74 percent year over year, and core (non-GAAP) revenue of $209.9 million, up 103 percent. The GAAP number carries a $450.5 million net loss, a figure that looks alarming until you see the core net loss for the same quarter: $6.9 million. The roughly $444 million gap is driven overwhelmingly by two items, per the company's own filing: a non-cash stock-based compensation charge tied to double-trigger RSUs that vested at the IPO, and OpenAI warrant amortization treated as a revenue offset under standard accounting rules. Neither is a cash outflow tied to the day-to-day business.

Cash, cash equivalents, restricted cash, and short-term investments totaled $8.6 billion as of June 30, 2026, with another $850 million in undrawn debt capacity. Remaining performance obligations, the company's contracted backlog, stood at $25.4 billion, tied to more than 600 megawatts of data center capacity under contract. Full-year 2026 core revenue guidance sits at $880 to $890 million.

Backlog Concentration Is Migrating, Not Disappearing

Cerebras has carried heavy customer concentration since well before its IPO. Early S-1 filings showed G42 alone at 83 to 87 percent of revenue in 2023 and 2024. The most recent full disclosure, in an April 2026 S-1 amendment, splits that concentration between MBZUAI at 62 percent and G42 at 24 percent, an 86 percent combined share, both entities linked to Abu Dhabi. The company's Q2 2026 filing describes a small number of customers each contributing 20 to 49 percent of revenue without naming a specific current split.

What's changed, per secondary analysis of the backlog composition that has not been independently confirmed against the underlying filing text, is that OpenAI now represents a large and growing share of that $25.4 billion figure. That's a trade, not a fix: less exposure to Gulf state counterparty risk, more exposure to a single, fast-moving AI lab whose own capital needs and priorities aren't fully public.

Q2 2026 Snapshot

Figure	Value	Source
GAAP revenue	$180.1M (+74% YoY)	10-Q / 8-K, filed 8/12/26
Core (non-GAAP) revenue	$209.9M (+103% YoY)	10-Q / 8-K, filed 8/12/26
GAAP net loss	$450.5M	10-Q / 8-K, filed 8/12/26
Core net loss	$6.9M	10-Q / 8-K, filed 8/12/26
Cash + short-term investments	$8.6B, plus $850M undrawn debt	8-K, filed 8/12/26
Remaining performance obligations	$25.4B, as of 6/30/26	10-Q, filed 8/12/26
FY2026 core revenue guidance	$880M – $890M	8-K, filed 8/12/26
Sept 1, 2026 close	$172.62 (-6.31%)	Aggregator (stockanalysis.com)

Filing figures are drawn from the company's own August 12, 2026 SEC filings as republished on Cerebras's investor relations site; the raw EDGAR documents could not be pulled directly due to a network access issue at time of writing. Price and sector-comparison figures are aggregator-sourced and not exchange-primary.

The Lockup Calendar Worth Circling

The bigger structural event isn't behind Cerebras, it's ahead of it. The main lockup, roughly 171 million shares, about five times the IPO's public float, releases at the earlier of two trading days after Q3 2026 earnings or 180 days after the May 14, 2026 prospectus date, which lands on approximately November 10, 2026. Q3 earnings haven't been scheduled yet, so whichever trigger comes first sets the date. That's the event with the real potential to move this stock on supply alone, not September 1's unexplained air pocket.

The Bull Case

• Growth is real and accelerating. Core revenue up 103 percent year over year and a $25.4 billion backlog against roughly $373.5 million in first-half 2026 revenue signal a company still early in converting contracted demand into recognized revenue.
• The GAAP loss is a distraction. Strip out the IPO-linked stock compensation charge and the OpenAI warrant amortization and the core business lost $6.9 million on $209.9 million of core revenue, a fundamentally different picture than the headline $450.5 million loss suggests.
• The balance sheet can absorb volatility. $8.6 billion in cash and short-term investments plus $850 million in undrawn debt is a deep enough cushion to ride out a rogue-wave session like September 1 without operational consequence.

The Bear Case

• Concentration risk didn't go away, it changed shape. Rotating from an 86 percent Abu Dhabi-linked customer base into heavy OpenAI exposure trades one single point of failure for another.
• GAAP gross margin is thin. The 14.2 percent GAAP gross margin in Q2, against a core margin closer to 41 percent that itself declined from roughly 46.5 percent in Q1, is worth watching for a trend, not a one-quarter blip.
• The stock has already shown it can move 15 to 20 percent on a headline. Between the June 24 earnings miss, the August 12 to 13 post-earnings drop, and the unexplained September 1 move, this name has posted four separate double-digit swings since its May IPO. The lockup unlock around November is the next scheduled one.

The SharkWater Take

I'm treating September 1 as noise until something surfaces to say otherwise. A move three times the sector's, with no filing behind it and a same-week reversal, reads like a liquidity air pocket in a stock that's been thinly seasoned since its May IPO, not a repricing of the business. The number I'd actually build a calendar around is the lockup unlock near November 10, roughly five times the current float coming free at once. That's a real, dated, mechanical supply event. Tuesday's six percent drop was not.

Tight lines. — SharkWater

Educational and informational purposes only. Not personalized investment advice. All figures sourced as noted and accurate as of publication. Options involve substantial risk of loss. The author may hold positions in securities discussed. Do your own work.