A good company and a good stock are not the same thing. I still think highly of Sterling Infrastructure (NASDAQ: STRL) as a business. I think the stock price has already absorbed a large part of the data-center growth still ahead of it.
That is why I am closing the position. The business thesis is not broken. The expected return no longer leaves much margin of safety. If the price opens that gap again, I will look at buying it back.
That distinction is the spine of this note. STRL’s business thesis is still alive. What I doubt is the valuation thesis.
This is the starting point for the physical-infrastructure leg of the AI Value Chain: compute only becomes usable capacity after power, land, and construction come together. The next piece, Applied Digital Corporation, follows the value of that scarce capacity through long-term leases and capital claims to common equity.
When something mattered more than a good product
While I was in the trading arm of Samsung C&T, I watched a neighboring team supply pipe, chemicals, and other materials from the early construction of Samsung Electronics’ fab in Taylor, Texas. That did not make me an expert on semiconductor plants. It did teach me one thing about how a large site actually moves.
Good products and good technology mattered. The question that stopped the job more often was different. Can it arrive quickly? Will it be there in the week the schedule needs it? Is there enough of it? One late component can push the work behind it. A supplier’s edge sits next to the spec: availability, timing, and the capacity to deliver that quantity.
This is not proof about Sterling. It is the intuition I am using. What matters more than the fact that a data center is needed is who can actually build it at the moment it is needed.
Sterling’s moat is closer to time than to technology
Sterling is not a wide technical moat. Reading it as a company with proprietary technology, network effects, software switching costs, or a licensed monopoly misreads the business. It is also not a semiconductor process barrier.
What it does have is a record of executing large sites, heavy equipment, skilled labor, project managers, electrical capability, customer relationships, the next phase of the same campus, and a footprint across the South, Northeast, Mid-Atlantic, and Rocky Mountain regions. CEC, acquired on September 1, 2025, added mission-critical electrical work to that list. The price was about $562 million, in cash and stock, plus an earnout of up to $80 million. CEC sits in E-Infrastructure. 2025 annual report
Money alone does not produce 2,000 skilled electricians, a heavy-equipment fleet, data-center site-development experience, and a project organization by tomorrow. When demand outruns supply, the firm that already has that capacity becomes more valuable. Sterling’s moat is closer to time than to technology.
Some of that time was accumulated. Some of it was bought. CEC was also a way to purchase electrical capacity and shorten the wait. A competitor can try the same kind of deal. A seller has to exist, the people have to stay, and the work has to hold its quality after the close. Capital does not finish those things on the day the money moves. I do not read this as a permanent monopoly. I read it as the ability to live through a few years of scarce supply while demand is sharp.
The second quarter of 2026 shows that shape. E-Infrastructure revenue rose 192% year over year to $905 million, 78% of the quarter. The company said the legacy site-development business grew revenue 111%. CEC’s electrical revenue rose 140% versus the pre-acquisition quarter a year earlier. Mission-critical work — data centers, manufacturing, and semiconductors — was 92% of E-Infrastructure backlog. Transportation revenue fell 20% in the same quarter. Management said it is moving people toward higher-margin E-Infrastructure work. Q2 2026 results
The growth is not only a proprietary technology story. Customers want data centers built, and the supply of firms that can grade the site and energize it has not kept up. Sterling stands on the short side of that gap.
The price moved. The business did not break.
The business did not collapse. The multiple did.
The month-end close was $142 in January 2025, $306 at the end of 2025, and $861 at the end of May 2026. On June 4 the stock traded as high as $1,005.68 and closed at $993.74. The October 6 close was $563.69, about 44% below the high. Guidance over that stretch does not look like a broken company. The February 2026 full-year revenue guide was $3.05–3.20 billion. On August 3 it was raised to $4.00–4.15 billion. February guidance, August guidance, prices
I am not claiming that the guide on June 4 was still the February range. First-quarter results sit in between. The fact I can check is simpler. Full-year guidance rose between February and August, and the stock fell from the June high to the October close.
Data centers are being added very fast
Mixing these numbers makes the market look larger than Sterling’s market. Total global data-center spending is not Sterling’s revenue pool. GPUs, servers, storage, and network equipment are not that pool either.
In an April 2025 article, McKinsey estimated that about $6.7 trillion of capital expenditure would be needed worldwide by 2030 for data centers to keep up with compute demand. Of that, $5.2 trillion was for AI data centers and $1.5 trillion for traditional IT. For AI alone, the three scenarios were $7.9 trillion accelerated, $5.2 trillion base, and $3.7 trillion constrained. Hardware is inside those totals. The cost of compute
Another McKinsey article says global data-center investment through 2030 is close to $7 trillion, more than $4 trillion of it is computing hardware, and more than 40% of the spending lands in the United States. It also says US demand could grow 20–25% a year over that period. The data center balance
Infrastructure excluding hardware is a third number. In August 2025 McKinsey said capital expenditure on data-center infrastructure excluding IT hardware would exceed $1.7 trillion globally by 2030. US power demand could rise from 25 GW in 2024 to more than 80 GW in 2030, more than tripling in about five years. Subtracting $4 trillion from $6.7 trillion and calling the result $1.7 trillion would glue different definitions into one table. They are not the same cut. Scaling bigger, faster, cheaper
US data-center power demand, 2024 and 2030
No years in between. This source does not publish them.
In October 2024 McKinsey estimated that capital spending on procurement and installation of mechanical and electrical systems for data centers could exceed $250 billion by 2030. That figure is global, and it mixes equipment makers with contractors. It is not a US-only construction market. It is useful because it shows that an electrical contractor does not sit inside a $6.7 trillion TAM. The same article said that even if known plans are delivered on time, the United States could still face a supply deficit of more than 15 GW by 2030. AI power
The more recent operating snapshot is CBRE’s first-half 2026 North America report. Supply in the eight primary markets rose 33.7% year over year to 10,903 MW. Capacity under construction rose 24.8% in the half to 7,481 MW, past the prior peak. More than 80% of it was already preleased. CBRE said power constraints are stretching completion dates. That reads less like weak demand and more like interconnection time pushing the build. CBRE H1 2026
Turner & Townsend’s 2025–26 construction cost index reported a 5.5% increase in the cost per watt of a traditional air-cooled data center. On US projects, liquid-cooled AI facilities of similar IT capacity typically cost 7–10% more to build than the air-cooled equivalent. The index is closer to construction cost excluding land and site works. Sterling’s site-development revenue is not fully inside that index. A large part of the electrical work is. Turner & Townsend
How much of that market can STRL actually take?
I will not pretend the addressable market is known to one significant digit. The range below is an author estimate.
Apply the “more than 40% in the United States” line from one McKinsey article to the $1.7 trillion ex-IT infrastructure pool from another, and US physical infrastructure lands around $0.7–0.85 trillion. That multiplication is not a fact from a single table. The 40% refers to total spending, including hardware. I am using it only as an analogy for the infrastructure pool.
Then take out power generation and transmission, land, the building shell, and mechanical cooling that Sterling does not do. If site development and mission-critical electrical work are about 25–45% of that US pool, the cumulative opportunity through 2030 is roughly $170–380 billion. The width is the point. It shows how many layers sit between trillions of total investment and this company’s revenue. It is not a precise market forecast.
From data-center spending down to Sterling
The upper rows are published estimates. Only the last row is the author’s. Do not add them.
If that range is about right, base-case 2030 E-Infrastructure revenue of about $6.4 billion is not a claim that Sterling owns the market. It is a business already above $3 billion, compounded at 20% for four years. Backlog points the same way. Signed backlog on June 30, 2026 was $4.33 billion, up 116% year over year and 50% organically. Combined backlog, including unsigned awards, was $5.62 billion, up 36% organically. The company said high-probability future phases exceed $1.4 billion, and signed work, unsigned awards, and future phases together exceed $7.0 billion of visibility. E-Infrastructure signed backlog rose 165%. Q2 results
Visibility is not 2030 revenue. Much of the backlog converts over 6–36 months, and future phases are not yet contracts. The 92% mission-critical share also includes manufacturing and semiconductor fabs, not only data centers. Reading 92% as a data-center market share would overstate the book.
In 2025 the top four E-Infrastructure customers were 27% of segment revenue, down from 31% and 40% in the prior two years. That is less concentrated than it was. It is still not a diversified software customer base. 2025 annual report
This is why the base case does not assume a large gain in market share. A bigger market is enough for the company to grow. The story in which Sterling keeps taking share as a construction platform is the bull case.
How much can it earn by 2030?
I did not build a detailed discounted-cash-flow model. I split the 2026 guidance midpoint, grow the pieces, apply a 2030 margin and an exit multiple, and discount that price to today.
At the August 3, 2026 guide, the midpoint is $4.075 billion of revenue, $903.5 million of adjusted EBITDA, and $20.00 of adjusted diluted EPS. The ranges are $4.00–4.15 billion, $891–916 million, and $19.70–20.30. The adjusted EBITDA margin on the midpoint is about 22.2%. Diluted shares in the company’s modeling notes are about 31.1 million. Guidance reconciliation
The company does not guide the segments for the full year. First-half E-Infrastructure revenue of $1.503 billion was 75.4% of first-half revenue of $1.994 billion. As an author assumption, I apply that mix to the full-year midpoint: $3.07 billion of E-Infrastructure and $1.00 billion for the rest. The rest is Transportation and Building. Management is already pulling people out of Transportation, so 0% growth in the rest is not an especially harsh case.
Growth runs four years from 2026 to 2030, not five.
| Scenario | E-Infrastructure CAGR | Other CAGR | 2030 EBITDA margin | Exit P/E |
|---|---|---|---|---|
| Bear | 10% | 0% | 18.5% | 18× |
| Base | 20% | 2% | 21.5% | 24× |
| Bull | 25% | 3% | 23% | 28× |
Those rows are not company guidance. They are author assumptions. The base margin of 21.5% is a little below the 2026 adjusted EBITDA margin of 22.2%. The base value does not depend on further margin expansion. The bear margin of 18.5% is normalization.
The bridge to adjusted EPS follows the company’s definition. Adjusted net income adds back stock-based compensation, intangible amortization, acquisition costs, and earnouts, tax-effected. On that definition, adjusted pretax income is roughly adjusted EBITDA minus depreciation of property and equipment. The 2026 midpoint of that depreciation is $73.5 million, 1.8% of revenue. I hold that ratio in 2030. I set net interest to zero. The 2026 midpoint includes $1–4 million of net interest income, and I do not treat that as a profit engine. The tax rate is 25%. The company describes an effective rate of about 24.5%, and putting this formula on the 2026 midpoint produces adjusted net income about 0.5% above the $621.5 million guidance midpoint. That gap does not change the conclusion, so I keep 25%. The share count stays at 31.1 million. No buybacks and no acquisition dilution.
Source: Company guidance; author estimates.
| Scenario | 2030 revenue | Adjusted EBITDA | Adjusted EPS | Exit P/E | 2030 price |
|---|---|---|---|---|---|
| Bear | $5.50B | $1.02B | $22.15 | 18× | $399 |
| Base | $7.46B | $1.60B | $35.41 | 24× | $850 |
| Bull | $8.63B | $1.98B | $44.10 | 28× | $1,235 |
Base-case revenue of $7.46 billion is $6.37 billion of E-Infrastructure plus $1.09 billion from the rest. The 2030 price is EPS times the exit multiple. $35.41 × 24 = $850.
A future $850 is not today’s $850
Put the 2030 price next to today’s price and the stock looks cheap. So I discount it. The required return is 10% a year. The clock runs from October 7, 2026, when this note is written, to December 31, 2030: 1,546 days, or 4.23 years. The discount factor is about 1.50.
| Scenario | 2030 price | Present value |
|---|---|---|
| Bear | $399 | $266 |
| Base | $850 | $568 |
| Bull | $1,235 | $825 |
| October 6 close | $564 |
What is STRL worth today?
2030 scenarios, brought back to today’s price
The base-case present value of $568 and the close of $563.69 sit inside the model’s noise. A 24.5% tax rate lifts the base present value to about $572. Cutting the full-year E-Infrastructure mix to 70% lowers it to $551. None of those turns the stock into something far cheaper, or far more expensive, than the base case.
The sentence I take from that is this. The current price is close to base-case fair value. The company did not get worse. A good deal of the future success is already in the price. I wrote separately about the gap between what is already priced and the next piece of news in a Bayesian look at efficient markets.
What does the price require?
Run the same math backward and the price states its demand.
For $563.69 to compound at 10% a year through the end of 2030, the stock has to be about $844 then. At the base-case exit multiple of 24, that requires adjusted EPS of about $35.2. From the 2026 midpoint of $20.00, that is 15% a year for four years.
At 18 times, the required EPS is about $47, close to 24% a year. At 28 times, it is about $30, or about 11% a year. A higher multiple asks less of earnings growth. If the multiple falls back toward an ordinary industrial level, earnings have to grow much faster.
The current price is not absurdly expensive. The problem is that the current price already requires a lot of success. If EPS grows about 15% a year and the exit multiple is 24, the expected return is about the 10% I require. There is no extra margin of safety left at this price.
What was inside the price near $1,000?
From the June 4 close of $994, a 10% compound annual return through the end of 2030 requires a price of about $1,540. That span is 1,671 days, or 4.58 years. At 24 times, EPS would need to be about $64, which is 34% a year from the 2026 figure of $20. Even at 28 times, EPS of about $55 is 29% a year.
I am not calling that price irrational. What I see in it is data-center demand plus a meaningful gain in share, a high margin, and a high terminal multiple, all at once. I cannot prove that combination is impossible. It does require many more things to go right together than the path I use as a base.
While the stock later fell to $564, the August full-year guide was higher than the February guide. That is why I judged a large part of the earlier valuation to have come from the AI-infrastructure narrative’s multiple, rather than from the improvement in the fundamentals alone. The business did not collapse. The multiple did.
Why I am selling STRL
I am not selling because I think the business is failing. Even if the base case is right, the expected return at this price is about the 10% I require. The bear present value of $266 is a long way down. The bull present value of $825 is not free. It needs 25% growth, a 23% margin, and 28 times, together.
When I size a position, I do not want a full weight at fair value. A good company with no margin of safety is not the same as a good company at a good price. If data-center spending, project delays, or the mood around AI capital expenditure push the price meaningfully below base value, I will look again. I will not lock that price in with today’s model. The assumptions have to be recalculated then.
Where I am most likely to be wrong
The weakest assumption in this note is that Sterling remains a fast-growing contractor and does not take a large amount of data-center construction share. If that assumption breaks, the bull case is closer to reality than the base case. At that point $564 may not look expensive.
The bullish risks to the sell are these. Data-center construction could run faster than 25% for longer. Cross-selling site work and CEC’s electrical work could be larger than I allow. More acquisitions could keep the growth rate up. The market could still pay 25–30 times after 2030. Buybacks would raise EPS on the same earnings.
The other side is open too. Hyperscaler capital expenditure can slow. Demand can stay strong while power interconnection and permitting delay the actual start and finish of construction. That delay is already in the CBRE note. Customer concentration, skilled-labor wages, competitors adding capacity, margin normalization, and the integration of CEC and Stone Ridge remain. The guidance includes the Stone Ridge acquisition. If integration is worse than planned, the 2026 midpoint itself can be too high.
Third-quarter results are not in this note. If the August guide moves, the 2030 starting point moves with it. The right response is to rebuild the table from the new guide, not to bend the conclusion toward the latest price.
I am not waiting for a good company to become a bad one. I am waiting for a good company to be a good price again.
Disclosure: This note records personal research and a personal investment decision. It is not a recommendation to buy or sell any security. At the time of writing, the author is reducing or exiting an STRL position, and that position may change without further notice.
Discussion
Discussion
Have a question, correction, counterexample, or another interpretation?