ERock promotes a $1.7bn backlog and its accountants signed off on $1.8bn — the first company in this section where the headline number is the conservative one

ERock assembles natural-gas reciprocating-engine generators and sells them, increasingly, to AI data centres that cannot wait for a grid connection. It listed on 11 June 2026 at $21.50 and traded as low as $8.88 seven weeks later. Then on 11 August it reported a quarter in which revenue fell 41.7% year over year and announced a backlog up roughly ten times, including a 470 MW equipment order from Anthropic — and the stock rose.

Both halves of that are real, and the interesting part is what happened when this report tried to take the bear side.

It lost, three times. The promoted backlog of $1.7bn turns out to be smaller than the audited figure: GAAP remaining performance obligations at 30 June were approximately $1.8bn. The full-year guide of $435–465m against $71.6m of first-half revenue looks like it needs a margin transformation; run the bridge and it needs volume, not margin. And the $(0.06) loss per share that beat a $(0.28) consensus is not mainly an Up-C attribution artifact — it covers 19 days.

What is left after all that is the thing worth watching. Customer prepayments rose $358.4m in six months to $528.4m, which is more than the entire second half the company has guided to. The money is in. The contracts are in. The question is a building in Houston.

What the market got wrong in ERock's favour

Market Storm has now written eleven of these, and ten of them found the same shape: the number in the press release is bigger than the number the filing supports. ERock is the first one that runs the other way, and it is worth being exact about it because the exception is more instructive than the rule.

The headline is "Contracted Power System Sales Backlog increased to approximately $1.7 billion, up 10x year-over-year." That is a management-defined operating metric — the registration statements describe it as "signed customer commitments with defined delivery schedules," it is not an ASC 606 disclosure, and it is not auditable line-for-line against the financial statements. All of that is a fair caveat and it is worth keeping.

But the GAAP figure sitting next to it, in the 10-Q's own remaining-performance-obligation note, is approximately $1.8 billion, expected to be recognised over "the next one to two years." The audited number is the larger one.

This report went looking for the opposite. The natural short thesis — and the one in circulation — held that GAAP RPO was $1.38bn against a $1.7bn promoted backlog, a ~$320m gap. That comparison came from an article published on 13 June 2026, two days after the IPO and two months before this quarter, and the RPO table it cites is dated 31 December 2025 and totals $339.5m. It is a June backlog measured against a December accounting table. On matched dates the sign inverts.

The two metrics genuinely do measure different things, and the differences run in both directions: RPO additionally carries multi-year ongoing-services obligations, which lifts it above a power-system-sales-only backlog; RPO also excludes usage-based variable consideration and contracts with an original duration of a year or less, which pulls it down. A spread of $100m either way between two differently-scoped metrics is bookkeeping. What it is not is evidence of promotion.

What the guide actually requires

Here is the arithmetic everyone starts from, and it is correct.

First-half revenue was $71,614k — Q1 $31,736k plus Q2 $39,878k — against full-year guidance of $435–465m. So the second half must produce $363.4m–$393.4m, roughly 84% of the year in two quarters, and about 5.1x to 5.5x the entire first half.

Three qualifications, all of which the arithmetic alone hides.

The base is worse than 5.1x, not better. Of that $71.6m, $29,178k is ongoing services — a recurring installed-base line running $13–16m a quarter that does not need to ramp at all. Strip it out and the burden falls entirely on power system sales: from $42,436k in the first half to roughly $330–365m in the second, which is closer to 8x.

But the comparison to "5.5x growth" is unfair in the other direction. The right like-for-like base is the prior second half, not the current first half. FY2025 revenue was $183.1m against first-half 2025 of $92.6m, so second-half 2025 was about $90.5m — making the implied year-over-year step about 4.0x–4.3x. And the "first half declined 22.6%" framing is entirely a Q2 artifact: Q1 2026 was up 31.6% year over year, and Q2 was up 26% sequentially and beat consensus. A single quarter of equipment-delivery timing is doing all the work in that percentage.

And the "per-quarter average" is a construct management contradicted. $182–197m per quarter is arithmetic, not guidance. Power system product revenue is recognised point-in-time on delivery, and the company guided to a ramp through year-end, with Hyperion reaching ~1.2 GW of annual assembly capacity only by then. Q3 will land well below $182m and Q4 well above $197m. The average understates the real Q4 requirement.

The margin question is the wrong question

The tidy bear line is that first-half adjusted EBITDA of $(26.9)m — a −37% margin — has to become +$29.9m to +$35.9m, an 8–9% margin, in six months.

Stated that way it sounds impossible. Run the bridge and it is mostly absorption.

That is 92% of the way there with no margin improvement whatsoever. Allow operating expense to grow a quarter, to $50m, and the required gross margin is about 22.0% — which is simply last year's 22.2%.

So the guide is a fixed-cost-absorption story, and the honest bear case has to attack delivery, not margin. That distinction is not pedantry: it tells you which line in the Q3 print falsifies the thesis. Watch power system product revenue. Do not watch gross margin.

One correction to carry: the widely-repeated first-half adjusted EBITDA of $(26.4)m — the sum of the two quarterly figures management spoke on the call — is $(26.9)m on the company's own six-month reconciliation. There is no Q1 10-Q; ERock listed on 11 June and this is its first periodic report, so the Q1 figure exists only as unaudited management commentary on a non-GAAP measure the company defines itself, with an "other items management deems non-operational" clause doing real work.

The $(0.06) that covers nineteen days

The reported loss per share was $(0.06) against a consensus near $(0.28), and it was widely written up as a 79% beat.

Consolidated net loss for the quarter was $67.7m.

The Up-C structure is the explanation people reach for, and it is partly right: continuing unitholders hold 79.95% of ER Holdings' economics, so most of the consolidated loss is allocated to non-controlling interests, which carry $207.3m on the balance sheet. But that is the smaller effect. At ERock, Inc.'s ~20.05% economic interest, a $(2.98)m attributed loss back-solves to a post-IPO consolidated loss of only about $14.9m — which means roughly $52.8m of the quarter's $67.7m sits in the pre-IPO period and never enters the Class A denominator at all. That includes essentially the entire $48,774k loss on debt extinguishment from converting notes and repaying a $30m term loan, both of which the filing places before the offering.

Two distortions stacked, and the period cutoff is by far the larger. The weighted share count tells the story on its own: 48,174,023 Class A shares, for a period running 11 to 30 June.

While correcting: $67.7m is a three-month figure, not a half-year one. Q1 added $(17.2)m, so first-half consolidated net loss is nearer $84.9m.

The deposits are the balance sheet

This is the part of the quarter that deserved the headline and did not get it.

Contract liabilities — customer prepayments on undelivered systems — went from $170,025k at 31 December to $528,405k at 30 June, a $358.4m build in six months. That single line is what produced first-half operating cash flow of $268.9m, against $0.4m in the comparable period.

Three things follow, and two of them cut against the company.

It is the strongest fact in the bull case. $528.4m of collected customer money exceeds the entire guided second half by $135–165m. Whatever else is uncertain here, the customers have committed cash, not just signatures. The CFO describes "significant upfront deposits… given market conditions," with milestone payments tied to long-lead equipment purchases and generator delivery.

It is also most of the cash. "Ended the quarter debt-free with $626.6m" is true of borrowings and misleading as a liquidity statement: 84% of that cash is an unperformed obligation, and a further $34.2m of the $660.9m headline is restricted as letter-of-credit collateral in favour of one major customer. Unrestricted cash net of contract liabilities is about $98.2m.

And the cash build is the IPO, not the deposits. Cash went from $108,097k to $626,636k — a rise of $518.5m against ~$554.0m of IPO net proceeds. Everything else combined, including $268.9m of deposit-driven operating inflow less capital spending, inventory and the ~$33m debt paydown, netted to a drain. Strip the deposits out of operating cash flow and first-half operations consumed roughly $90m: accounts receivable rose $68.0m to $101,790k and prepaid expenses $13.8m, together absorbing $89.5m of the inflow, with inventory at $106,059k pre-built for the ramp.

One timing caveat this report could not settle: a single unconfirmed read of the pre-IPO balance sheet put contract liabilities near $471m at 31 March, which would place ~84% of the deposit build in Q1 — before the Anthropic order was disclosed — and make Q2's sequential addition only ~$57m. That figure could not be corroborated and is flagged, not used.

Whether the megawatts are physically real

The product is the RockBlock: reciprocating natural-gas engine generator strings scaling in 0.5 MW increments from 1.5 to 3.5 MW per unit, assembled in-house. Not turbines, not fuel cells. Its differentiator is transient response — full load acceptance in as little as 10 seconds against up to 45 for competing gas solutions — and a diesel-replacement footprint, rather than thermal efficiency, where Bloom's fuel cells at ~60% beat it comfortably.

"Proprietary" is doing work in that sentence. ERock assembles; the S-1's own risk factors describe suppliers that are "limited and in some cases sole-sourced," using proprietary processes, with no comparable alternative "without considerable delay." The company's engine lineage runs through a long-term mutual-exclusivity supply agreement with Power Solutions International for gensets built on PSI's 21.9-litre natural-gas heavy-duty engine. Nameplate assembly capacity is not the same as deliverable megawatts when upstream components are allocated in the same 2028-constrained market as everyone else's — Cummins data-centre allocation reported sold out through 2028, Caterpillar backlog around $63bn, industry lead times near 42 weeks.

The economics work and are improving. Modern medium-speed simple-cycle recips run ~8,000–8,400 Btu/kWh HHV at $1,250–1,650/kW installed with >96% availability. Thunder Said models a 5 MW unit needing 8–10¢/kWh for a 10% IRR at 60% utilisation — above ERCOT North's $47.39/MWh baseline forecast, but EIA's high-demand case has that hub up ~79% by 2027, and at near-zero Waha gas with baseload utilisation the case is far stronger than a 60%-utilisation model implies. Fuel is a tailwind: Waha traded negative for 118 of the first 131 days of 2026, at times near −$10/MMBtu, with 44.9 Bcf/d of new US pipeline capacity planned for 2026–27, two-thirds of it originating in Texas.

Permitting is manageable rather than blocking. TCEQ's Air Quality Standard Permit for Natural Gas Electric Generating Units took effect on 30 January 2025 and lets many gas EGUs skip case-by-case new-source review; TCEQ issued a 7.65 GW West Texas air permit in 2026. The scale is real — a recip-engine data centre in Tom Green County applied with 419.3 tons/yr of projected NOx — but Texas will permit this class of plant.

The policy event that is both cases at once

The sharpest thing in this research is that one directive is simultaneously the bull case and the bear case, and I did not expect that to be true of a single event.

On 3 August 2026 — eight days before this 10-Q — the Governor of Texas directed the PUCT and ERCOT to verify and audit every data-centre project in the interconnection queue before any additional projects advance, and said non-compliant projects should be denied grid access. ERCOT tracked 474.7 GW of large-load requests as of June 2026, 420.8 GW of it data centres, and responded the same day that it would miss its 7 August Batch Zero deadline. BNEF estimates the audit could delay 49.8 GW of data-centre load and cost projects up to $15bn.

That is the single strongest demand driver imaginable for on-site generation. A customer served behind the meter is not waiting in a frozen queue — which is precisely why a backlog went up ten times.

And the audit explicitly examines on-site generation. SB 6, effective June 2025 with PUCT implementation due by 31 December 2026, covers loads at or above 75 MW, mandates curtailment protocols for large loads interconnecting after 31 December 2025, and imposes reliability studies and PUCT approval on co-location arrangements with behind-the-meter generation. The PUCT has already affirmed that emergency curtailment of a co-located data centre is not capped by the capacity of its paired behind-the-meter generator.

So behind the meter reduces exposure to the queue; it does not exempt the customer from Texas policy. And with over 80% of ERock's sales in Texas, the concentration and the tailwind are the same fact.

What the price assumes

At the 25 August close of $13.11, on 219,400,080 fully-exchanged shares — 48,174,023 Class A plus 171,226,057 Class B as of 7 August — market capitalisation is about $2.876bn, and with $626.6m of cash against no borrowings, enterprise value is roughly $2.25bn.

State a divergence rather than average it: stockanalysis.com reports 273.30m shares and a $4.16bn market cap — about 54m shares and $1.28bn above the 10-Q cover count — which is the sort of Up-C share-count error that puts a screen's multiple 60% too high. This report uses the filing's count.

On that basis:

Against retrievable comparables — GE Vernova at 5.9x EV/revenue, Powell Industries at about 7.6x EV/sales, Bloom Energy at roughly 22x price/sales — ERock at 5.0x on guided revenue screens genuinely cheap. On trailing revenue it is 13.9x, or 21.2x on GuruFocus's higher enterprise value against a stated ~3.5x peer median.

The entire valuation argument is the denominator. At 5.0x the stock is cheaper than GE Vernova; at trailing revenue it is four times the peer median. Which of those is the real multiple is decided by whether the second half ships.

Two structural facts sit under the price. About a third of the ~$554m of net IPO proceeds went to pre-IPO holders — roughly $156.9m purchasing Class B units and $27.8m of blocker-merger cash, including ~$119.9m to sponsor-affiliated funds — so "money raised" and "money into the business" differ materially. And the tradeable float is only about 19m shares, smaller than the 27.9m-share IPO itself, with short interest of 5.7m shares reading as 29.9% of float or 2.08% of shares outstanding depending on which base you use. The 39% de-rate from $21.50 happened in a stock where almost nothing is actually tradeable. That is a thin-float price, not a considered one — and 171.2m Class B units become exchangeable around mid-December.

Worth noting against the bear case: the stock is up since the print, from $11.25 on 11 August to $13.11 on 25 August, and every analyst action afterwards was a raise or a bullish initiation — Morgan Stanley $21→$23, BofA $16→$19, Evercore ISI initiating at $28, Barclays at Overweight. No target was cut. Mean target $22.88, low $19, 8 buys and 0 sells.

Risk — each isolated, do not blur

1. Delivery (dominant). Power system sales must go from $42.4m in the first half to roughly $330–365m in the second, out of a facility that began operating in Q2 at a company with, by its own admission, a limited history of assembling its own systems.
2. Supply chain. The engine is assembled around sole-sourced components in a market where large-genset slots are allocated into 2028. Assembly nameplate does not equal deliverable megawatts.
3. Concentration. Over 80% of sales in Texas; roughly half of 2025 sales from three clients; backlog growth dominated by one 470 MW order.
4. Backlog quality. The Anthropic order is equipment with O&M contracted separately — one-time hardware margin, not the recurring microgrid annuity the pre-IPO business was built on.
5. Texas policy. The August audit and SB 6 cut both ways, and the PUCT has ruled that a paired behind-the-meter generator does not cap curtailment of a co-located load.
6. Supply of stock. ~19m float, 171.2m units unlocking around mid-December, a TRA paying 85% of net cash tax savings to pre-IPO holders, and 29.9%-of-float short interest — all landing in the same weeks that Q4 delivery resolves.

Horizon and sizing (kept separate)

Horizon. Unusually legible and unusually near. The Q3 print is where the arithmetic becomes falsifiable: if Q3 power system revenue is not materially above $100m, the full-year guide is arithmetically dead regardless of backlog. Watch power system product revenue and the contract-liability balance — a deposit balance that keeps building while revenue does not is the signal that delivery, not demand, is the constraint.

Sizing considerations (not a recommendation). This is a ten-week-old listing with one periodic filing, a ~19m-share float, a lock-up expiry inside four months, and a full-year outcome concentrated in a single quarter of factory output. Whatever the multiple says, the position sizing that matters here is governed by liquidity and event concentration rather than by valuation. It is also the second report in this section — after IREN — where the company owns the physical bottleneck rather than renting it, which is a genuinely different risk profile from the neoclouds and correlates with them less than the ticker tape currently suggests.