Comfort Systems USA

The tagged GAAP segment operating margins run the other way from the gross margins: Mechanical at 18.6% and Electrical at 15.9% in the same quarter, so the Electrical segment earns the better gross margin and the worse operating margin, and carries more overhead per dollar. More importantly, the chief financial officer disclosed on the same day that first-quarter gross margin included gains he called 'outside the scope of the ordinary', so the sequential comparison the release invites - 26.3% down to 25.9% - is not the like-for-like one; his own restated comparison is 25.2% up to 25.9%.

Every one of these multiples is struck against a company whose return on invested capital is 38.8% against a peer median of 17.9% and its own six-year median of 19.5%, and whose trailing revenue grew 46.1% against a peer median of 18.9%. A multiple that has roughly doubled against a return on capital that has roughly doubled is not prima facie irrational, and the price-to-sales ratio at 4.0 times its own history overstates the change because the margin on those sales also doubled. The exception cuts the other way rather than defending the price: the one metric on which Comfort Systems looks reasonable is the one struck on a cash flow its chief financial officer has said reverts to net income.

The float is a genuine competitive fact and not merely an accounting one - the president reads it as evidence that customers 'value locking up our capacity right now' - and a contractor whose customers prepay is in a stronger position than one financing its own work in progress, which the $159.3 million reverse balance shows Comfort Systems barely does. The qualification is what it does to the return figures rather than to the business: a 38.8% return on invested capital is struck on a capital base that customers have very largely funded, and the increase in the float over nine months is 69% of the operating cash flow for those nine months. It is not a reserve, it is work owed.

This is management telling the market not to extrapolate its best number, which is candour rather than weakness, and the same executive noted the company has converted above net income 'for years'. But the guidance is explicit and it is the one forward statement that bears directly on the valuation: the free-cash-flow multiple of 26.2 is the only metric on which this company is not the dearest of its filed peers, and it is struck on a cash flow its own chief financial officer has said will fall back to roughly net income. On trailing earnings the same price is 39.6 times.

The return figures are the strongest evidence in the file and they carry the qualification in C10: invested capital is small because customers funded it. Billings in excess of costs of $3,231,104 thousand is almost exactly equal to total stockholders' equity of $3,217,269 thousand, so a dollar of customer prepayment sits behind every dollar of book equity. That does not make the return false - the cash is real and it is deployed - but it does mean the denominator shrinks when the float stops growing, and the return is a function of the growth rate as much as of the margin.

The chief financial officer's defence is substantive rather than evasive and deserves weight: a percentage-of-completion contractor that never booked net favourable revisions would be systematically under-accruing risk, and he says the revisions have been positive every year for two decades. He also gives a mechanical reason for the increase - 'the number of jobs we have over $40 million or $100 million counsels us to not rush to recognize revenue on things that aren't finished', so bigger jobs mean more conservatism at the start and larger releases at the end. Against that: the contribution rose every year of the annual series, from 1.3% to 2.3% to 3.9%, and to roughly 7.7% in the June quarter. The September quarter's 4.5% is the first fall, back near a year earlier's 4.0%, and the nine-month 6.6% is still above any full year.

The discipline is real and is the opposite of the speculative capacity race that usually ends a construction cycle: capital goes in only against commitments, and the chief financial officer says paybacks have been 'within a year or two'. The cost is concentration. The expansion is entirely for existing customers on the president's own account, so the capital being spent deepens a customer base already carrying one name at 12.8% of revenue and one sector at 58.7%, and the diversification that would offset it exists only as pilot contracts nobody has sized.

The maintenance base is a contract-value figure and service revenue is larger than it: the same presentation puts service and maintenance at 6% of year-to-date revenue, which on a $12 billion annualised run rate is roughly $720 million, because a maintenance agreement generates call-out and time-and-materials work beyond its contracted value. The base is also not the only thing a second source cannot take - existing-building construction was 15% of year-to-date revenue and is performed inside occupied buildings, and every modular module still has to be craned, connected and commissioned on site by licensed trades. The finding is one of size rather than of direction: the portion of this business that renews without being re-won each year is $185 million, and it is growing at about a thirtieth of the rate of the portion that must be.

Backlog is not contracted revenue. The company's own risk factors state that 'our backlog is subject to unexpected adjustments, cancellations and scope reductions and, therefore, may not be a reliable indicator of our future earnings', and separately that 'projects are cancelled, delayed or modified due to customer, industry or macroeconomic conditions'. Backlog of $17.81 billion is now about 1.4 times trailing twelve-month revenue of about $12.5 billion; the change is the growth rate rather than the coverage, and four operations account for the whole sequential rise.

Lane has since sold 11,113 shares in May and 16,024 in August 2026, outside any trading plan, leaving 145,065 (C22), so the largest insider stake is smaller than the proxy shows.

The 2025 targets sat close to 2024's results (EPS of $14.60, C02), so a maximum payout says more about how low the bar was set in a boom than about stretch; nothing in the plan pays on return on capital or on backlog quality, and EPS rewards the estimate revisions C08 records.

Officers and directors of a company whose shares have risen 133% in a year sell for reasons that have nothing to do with a view, and none of these filings names one. The discretionary point cuts both ways: a company with no 10b5-1 plans in place has no scheduled selling either, so its insiders can only ever sell in an open window, and the August sales cluster in the window after the July results exactly as compensation-driven selling would. What makes the pattern worth recording is the asymmetry rather than any one trade: 45 transactions, 30 open-market sales and not a single open-market purchase, through a 20% fall from the high that the company's own results have given no reason for. Nobody inside bought the decline, and the chief executive sold a tenth of his stake into it two days after the reference price this thesis was first struck at.

A risk-factor section is written to disclose risk and will always describe an industry at its least flattering, so this is not a neutral account. The offsetting evidence is that the description has not changed while the margin has: the same low-barrier language appeared in prior years, and gross margin still went from 19.0% to 25.9%. That gap is what the moat block has to explain, and it is why the mechanism is recorded as scale showing up in pricing rather than as anything structural.

A gross margin that rises 690 basis points in ten quarters in an industry the same filing calls price-competitive with low barriers to entry deserves more suspicion than the company invites. Management attributes it to volume leverage and 'improved execution in our operations', and the largest single named contributor in 2025 was $124.3 million of 'improvements in project execution at our Texas modular operation' - which is real, and is also the operation whose estimates C08 shows are being revised upward at an accelerating rate.

The tagged GAAP segment operating margins run the other way from the gross margins: Mechanical at 18.6% and Electrical at 15.9% in the same quarter, so the Electrical segment earns the better gross margin and the worse operating margin, and carries more overhead per dollar. More importantly, the chief financial officer disclosed on the same day that first-quarter gross margin included gains he called 'outside the scope of the ordinary', so the sequential comparison the release invites - 26.3% down to 25.9% - is not the like-for-like one; his own restated comparison is 25.2% up to 25.9%.

The return figures are the strongest evidence in the file and they carry the qualification in C10: invested capital is small because customers funded it. Billings in excess of costs of $3,231,104 thousand is almost exactly equal to total stockholders' equity of $3,217,269 thousand, so a dollar of customer prepayment sits behind every dollar of book equity. That does not make the return false - the cash is real and it is deployed - but it does mean the denominator shrinks when the float stops growing, and the return is a function of the growth rate as much as of the margin.

EMCOR is describing itself as much as the industry, and a competitor's claim to advantage is no more disinterested than the subject's. What makes it evidence is the agreement across two adversarial filings on the point that hurts both: neither claims a structural barrier, and both locate their edge in the same four places. The reading that cuts the other way is scale-dependent - EMCOR is nearly twice the size on revenue and earns 12.8% at the segment level against Comfort Systems' 17.1% at the company level, so if scale alone conferred the margin the ranking would run the other way. Something other than size is doing the work, and the candidate is the modular mix.

The benign reading is management's own and is plausible on the numbers: if customers are seeding second sources because demand exceeds what Comfort Systems will build, that is a statement about the size of the market rather than about the company's position in it, and it is consistent with a backlog up 73%. The uncomfortable half is in the same sentence and management does not dismiss it either: the product was 'co-designed with our customers', so the design does not belong to Comfort Systems, and the buyer is choosing who else builds it. A capability the customer can transfer is a lead, which is what the executive calls it, rather than a moat.

Value put in place measures construction spending, not orders, so it records the build that has already been committed and is a lagging read on demand rather than a forward one - it would keep rising for some quarters after hyperscaler capital plans were cut. It is not seasonally adjusted at this level of detail, which is why the comparison is year-to-date against year-to-date rather than month against month, and the single-month figure carries a preliminary marker. The revision to June - up $304 million on the figure quoted in the first pass of this thesis - is a reminder that recent months move. The offsetting point is that this is exactly the property that makes the series useful against a company statement: it measures work actually done, by an agency with no position in the outcome, and it is the same series read twice sixteen days apart rather than a new one chosen because it agreed.

Backlog is not contracted revenue. The company's own risk factors state that 'our backlog is subject to unexpected adjustments, cancellations and scope reductions and, therefore, may not be a reliable indicator of our future earnings', and separately that 'projects are cancelled, delayed or modified due to customer, industry or macroeconomic conditions'. Backlog of $17.81 billion is now about 1.4 times trailing twelve-month revenue of about $12.5 billion; the change is the growth rate rather than the coverage, and four operations account for the whole sequential rise.

At 12.8% this is disclosed concentration rather than dependence, and it sits below the thresholds that would make a single customer existential. The reason it matters is the direction of travel around it: technology customers went from 45.0% of revenue in 2025 to 58.7% in the June 2026 quarter while this disclosure was being written, and the modular capacity expansion is committed to existing customers, so the 2026 figure is very unlikely to be lower. The company does not disclose customer concentration quarterly, so the next reading is a year away.

The discipline is real and is the opposite of the speculative capacity race that usually ends a construction cycle: capital goes in only against commitments, and the chief financial officer says paybacks have been 'within a year or two'. The cost is concentration. The expansion is entirely for existing customers on the president's own account, so the capital being spent deepens a customer base already carrying one name at 12.8% of revenue and one sector at 58.7%, and the diversification that would offset it exists only as pilot contracts nobody has sized.

Value put in place measures construction spending, not orders, so it records the build that has already been committed and is a lagging read on demand rather than a forward one - it would keep rising for some quarters after hyperscaler capital plans were cut. It is not seasonally adjusted at this level of detail, which is why the comparison is year-to-date against year-to-date rather than month against month, and the single-month figure carries a preliminary marker. The revision to June - up $304 million on the figure quoted in the first pass of this thesis - is a reminder that recent months move. The offsetting point is that this is exactly the property that makes the series useful against a company statement: it measures work actually done, by an agency with no position in the outcome, and it is the same series read twice sixteen days apart rather than a new one chosen because it agreed.

The series is noisy month to month - it printed 1.6% in December 2024 and 4.8% eleven months earlier - so a five-month run is a short trend, and it is measured over all of construction rather than over the mechanical and electrical trades this company competes for. The level also does not support the strongest version of the scarcity claim: at 3.8% the market is tighter than at any point since August 2024 and materially looser than in 2018, 2019 and 2022, so a reader told that labour is the binding constraint on the industry should know the constraint is not at a record. What the series does establish, and the company's own assertion cannot, is direction: this is tightening while the company is pricing as though it were.

'Technology' is the company's own customer-type label and is broader than data centres - it takes in semiconductor fabrication and other high-technology manufacturing plants, and the company does not disaggregate it further. So the 58.7% figure is an upper bound on data-centre exposure rather than a measure of it, and the prize block treats it as a bound for exactly that reason. The direction is not in doubt: the 10-Q attributes the same-store increase specifically to data centres, and names the individual operations - Texas modular, Texas electrical, North Carolina, Indiana - where it landed.

At 12.8% this is disclosed concentration rather than dependence, and it sits below the thresholds that would make a single customer existential. The reason it matters is the direction of travel around it: technology customers went from 45.0% of revenue in 2025 to 58.7% in the June 2026 quarter while this disclosure was being written, and the modular capacity expansion is committed to existing customers, so the 2026 figure is very unlikely to be lower. The company does not disclose customer concentration quarterly, so the next reading is a year away.

The benign reading is management's own and is plausible on the numbers: if customers are seeding second sources because demand exceeds what Comfort Systems will build, that is a statement about the size of the market rather than about the company's position in it, and it is consistent with a backlog up 73%. The uncomfortable half is in the same sentence and management does not dismiss it either: the product was 'co-designed with our customers', so the design does not belong to Comfort Systems, and the buyer is choosing who else builds it. A capability the customer can transfer is a lead, which is what the executive calls it, rather than a moat.

The chief financial officer's defence is substantive rather than evasive and deserves weight: a percentage-of-completion contractor that never booked net favourable revisions would be systematically under-accruing risk, and he says the revisions have been positive every year for two decades. He also gives a mechanical reason for the increase - 'the number of jobs we have over $40 million or $100 million counsels us to not rush to recognize revenue on things that aren't finished', so bigger jobs mean more conservatism at the start and larger releases at the end. Against that: the contribution rose every year of the annual series, from 1.3% to 2.3% to 3.9%, and to roughly 7.7% in the June quarter. The September quarter's 4.5% is the first fall, back near a year earlier's 4.0%, and the nine-month 6.6% is still above any full year.

The company is an installer rather than an importer: it buys equipment from domestic manufacturers, and the duty reaches it only through what those manufacturers pass on. Its own assessment is that costs are 'in many cases' recoverable and immaterial, and it has bid and delivered a 690-basis-point gross margin expansion through eighteen months in which metals tariffs were already rising, which is evidence against the alarming reading rather than for it. Against that: 'in many cases' is not all cases, the sentence is unquantified and identical in form to the one in the prior-year filing, and the mechanism bites specifically on fixed-price work bid before April 2026 that is still burning off a $14.06 billion backlog. Under the cost-to-cost method an unrecovered input increase lands as a downward revision to estimated profit on those jobs - the exact reverse of the upward revisions running at roughly 7.7% of revenue. Nothing here says that will happen; what it says is that a named mechanism moved, the only party who can size it has not, and the thesis assumed pass-through without writing the assumption down.