a16z says AI accounts for nearly all new construction spending, and Census shows data centers up 57.2% as the total fell 3.8%

What the report says

The deck puts hyperscaler capital spending at $416 billion in 2025 and $791 billion in 2026

The deck is written for investors, so its construction material arrives as evidence for a market argument. Capital spending by the largest cloud operators, which a16z calls hyperscalers and lists as Alphabet, Amazon, Meta, Microsoft and Oracle, was $241 billion in 2024 and $416 billion in 2025, according to J.P. Morgan Asset Management figures in the deck. Analysts’ consensus estimates put 2026 at $791 billion and 2027 at $1,066 billion. a16z writes that each year’s consensus forecast has come in below what the companies then spent.

That spending has cut the companies’ free cash flow sharply (free cash flow is the cash left after operating costs and capital spending). a16z expects the drop to last until about 2028, and it shows bond sales and off-balance-sheet vehicles covering part of the difference.

Goldman Sachs, cited in the deck, estimates that construction employment in data center-exposed categories rose by more than 300,000. The categories are HVAC contractors, electrical contractors, utility construction and commercial building construction. A slide on supply bottlenecks lists megawatt switchgear at 52 to 104 weeks, in a power and grid group that also names transformers and gas turbines.

What Census shows

Data centers added $27.4 billion in a year while every other category combined lost $112.4 billion

Census splits private office construction into general office, financial and data centers, and the data center line is where the growth is. It more than doubled in two years, from $36.1 billion SAAR in July 2024 to $75.2 billion in July 2026. It rose 6.2% from June alone.

In dollars, data centers added $27.4 billion from July 2025 to July 2026. Every other category combined lost $112.4 billion, which is how the national total fell $85.1 billion. So the a16z line holds for the categories that grew. Data centers posted the largest dollar gain of any detailed category Census reports, ahead of private electric power at $10.2 billion and highway and street at $6.5 billion.

The deck’s other construction theme, a return to building physical things, reads differently in the same tables. Private spending on computer, electronic and electrical plants, the category that holds chip fabs, fell 46.9% in a year to $51.3 billion SAAR, 59.4% below its peak of $126.4 billion in June 2024. Data centers passed it in April 2026, and they led it by $23.9 billion in July.

What BLS shows

Four data center trades added 109,400 jobs in a year while the rest of construction lost 33,400

BLS payroll counts back the labor half of the claim. They run a month behind for detailed industries. The four industries Goldman names employed 3,920,200 people in July 2026, a preliminary figure, which is 109,400 more than a year earlier. Total construction employment rose 76,000 over the same months, so the rest of the industry lost 33,400 jobs.

From January 2023, the four industries added 337,100 jobs, consistent with Goldman’s figure. The starting month of Goldman’s count is not shown in the deck. (These four industries also build hospitals, warehouses and plants, so the BLS counts include work beyond data centers.)

Pay and lead times

Superintendent postings for data center work showed $139,000 against $110,000 elsewhere

Indeed’s figures in the deck compare median posted base pay for the same job title on data center postings and on other US postings from January to June 2026. The superintendent difference is 26.4%. Construction manager postings showed $134,000 against $104,000. Project managers had a smaller gap, $115,000 against $103,000.

Indeed’s own July analysis shows how concentrated the hiring is. Data center postings have more than doubled in two years to about 6 of every 1,000 US job postings, and the 10 largest tech firms account for 71% of them in 2026. The postings cluster in a few places, including the outskirts of Columbus, Ohio; Jackson, Mississippi; Reno, Nevada; and Hermiston, Oregon.

Switchgear ordered in October 2026 would arrive between October 2027 and October 2028

For contractors

How to use this on your jobs

Anyway, the part of the deck a contractor can act on is the concentration it describes. The growth in construction spending and in trade jobs comes from a small group of owners, and a16z shows those owners borrowing to build while their free cash flow drops. These steps use records most firms already keep:

1. Tag each job in the backlog report by end market (data center, power, other private, public) and by owner. Then compute the data center share of backlog: contract value remaining on data center and power jobs divided by total contract value remaining. Capture it for the last four quarter-ends as a baseline.
2. On data center subcontracts, read the pay-when-paid clause, retainage terms and lien notice deadlines with the owners’ cash position in mind.
3. Track superintendent and project manager turnover by metro, and log the reasons candidates give for declining offers. Compare them with the pay gap Indeed reports, and price retention into bids in the metros Indeed names.
4. Release long-lead electrical gear at award, and write escalation language for it into the subcontract.

The deck also cites Apollo figures that apply to any AI pilot a contractor runs. In the second quarter of 2026, 69% of S&P 500 companies pointed to a live AI deployment and 2% disclosed a metric they track over time. Before a pilot starts, pick its metric, such as RFI turnaround in calendar days, and record the baseline on one job.

Before you use these numbers

Revisions and caveats

Census’s July figures are preliminary. Its average revision from the preliminary estimate to the first revision is 0.65% for total construction, and August figures come out October 1, 2026. The BLS counts for July in the four detailed industries are marked preliminary as well.

The a16z capex figures for 2026 and later are consensus estimates, and a16z says in its disclosures that it has not independently verified the third-party data in the deck. Indeed’s pay figures are median posted base pay on postings that disclose pay, so they measure advertised offers. Indeed counts a posting as data center-related when the term appears anywhere in its title or description, and Indeed notes that this picks up some roles unrelated to data centers.