The boom
this week

Sixteen gigawatts, signed through 2028

Two weeks ago Broadcom's CEO, Hock Tan, told analysts on the company's earnings call that it now has roughly 16 gigawatts of committed AI chip deployments for two customers through 2028. One of them, Anthropic, goes from 1 gigawatt this year to 5 next year to 10 in 2028, which would make it Broadcom's largest custom-chip customer. The other, OpenAI, has a 1.3-gigawatt program landing in 2027 with a line of sight to more than 5. Tan said demand is running ahead of Broadcom's own forecast, and the thing holding it back is fab capacity, not orders. (CNBC and others reported the call; the numbers are Tan's.)

Here's why a chip company's order book matters to a welder. A gigawatt of chips is not a warehouse of boxes. It's a gigawatt of buildings, substations, transmission, cooling, and gas turbines that don't exist yet, on a contract with a date on it. When a customer signs for 10 gigawatts in 2028, somebody has to pour the slabs in 2026 and hang the switchgear in 2027. These are not press-release numbers. They're purchase orders. That's the difference between a boom that could stop tomorrow and one that has a floor under it for the next two years.

One supporting number, with the caveat it deserves: Exponential View estimates global AI revenue is running at about a $229 billion annualized rate, roughly 3.5x in a year. Those are analyst estimates, and the most recent months are projections. But real revenue at that scale needs real buildings, and real buildings need the people on this board.

Two builds worth knowing by name

Grimes County, Texas. SpaceX is putting a $16.8 billion chip plant, the first phase of what it's calling Terafab, in a county of about 30,000 people an hour and a half north of Houston. Construction starts this year, phase one finishes in 2028, 3,000 jobs when it's running. The campus is planned at 100 million square feet. A semiconductor fab is the most trade-intensive building there is: electricians, pipefitters, sheet metal, welders, and millwrights for two years before a single engineer badges in. The 3,000 is what the state counted at the end; the construction crews come first and aren't in that number.

Williams County, North Dakota. My county. Valley News Live tracked every data center in the state last month: 26 facilities, almost 4,000 megawatts. Williams County leads with 10 facilities and 1,710 megawatts, close to half the tracked total. The next one is already approved north of Williston at 56th and Highway 2: Critical Data House, 455 megawatts, up to $5 billion, seven buildings, phased past 2027. Those are numbers you usually hear about Texas and Virginia.

The metal side

The chips get the headlines. The copper has to come out of the ground first, and the mines can't find the hands.

Copper World, Arizona. Hudbay's mine south of Tucson is the first fully permitted new copper mine of its size in the U.S. in years. Mitsubishi paid $600 million for 30% of it, closed this January. The plan is 85,000 metric tons of copper a year for 20-plus years, 400 or more direct jobs, up to 3,000 across Arizona. Seven building-trades unions have already signed the construction agreement. The final go decision is expected this year.

Nevada. The state's economic development office says mining pays an average of about $120,000 a year and it still can't fill the jobs, because the mines are in Elko and Winnemucca and the recruits are in Vegas and Reno and won't move. Officials figure about half the state's miners retire by 2030.

Alaska. The Alaska Miners Association's commissioned research puts the industry average at $122,568, nearly twice the state's private-sector average, across roughly 5,900 direct jobs. The board has been carrying Alaska rotation work all summer, fly-in fly-out.

Same story three ways: the money is there, the shortage is real, and it goes to whoever is willing to live where the ore is.

The rail

Three weeks after we moved the API to Solana, it's the top chain by app revenue by a factor of about two, and Robinhood's new chain just passed Ethereum for second place. That's DefiLlama's 24-hour snapshot from last week, and a 24-hour number is noisy, so read it as direction, not a scoreboard. The direction is that the agents are settling where it's cheap.

See the data API →

The board

8,171 listings this week, the biggest board yet. Be careful with that number. The aggregator feeds that ran thin the last two weeks came back hard, so the jump from 5,863 is mostly sampling, not hiring. Compare the trades, not the total: electrician 288, lineman 141, welder 281, HVAC 339, mechanic 265, CDL 243. Per diem 207, travel paid 60, sign-on 163 (top one $35,000), will-train 322, fair-chance 160, housing 35. North Dakota 83, up from 72. Texas 1,291.

Oilfield held flat. Raw supply up a little, board count down a little after the scrub, ND rows up, rig count 34. Not the continued slide we saw two weeks ago and not a clear reversal.

Two changes under the hood, because you're paying for honesty. First, the wage floor by trade now publishes the sample size next to the number, and if fewer than five employers posted real pay for a trade, the floor is null instead of a guess. This week lineman is null for the second week running: of 141 listings, fewer than five carry employer-posted pay. Linemen don't post the number. They call you. Second, the scrub got tougher: interns, sales, accounting, HR, legal, software, and support-desk roles are out, and a front-end loader operator no longer gets killed by the pattern that catches front-end developers. About 500 desk rows came off this week's board before you saw it.

The enriched API is live with the full signal set: rig count, EIA production and prices, grid demand, wage floors with sample sizes, top employers. Same address, a cent a call.

Next week

Sunday pull, Issue #11. If you got hired off this board, tell me. That's the only metric I actually care about.