Executive Summary
Today's sources point the same way. Producing software is getting cheap fast, and value is moving to whoever defines the problem, checks the result, and owns the outcome. SemiAnalysis projects about +50GW of datacenter capacity coming online in 2027, roughly the entire global installed base through 2024. That supply makes artifacts cheaper still. Scheff's lament (via Willison) and Nate Jones's "don't keep shipping" heuristic describe the human side: implementation scarcity no longer gives engineers leverage, and output volume no longer signals contribution. Enterprises that keep measuring PRs, tickets, and features will get more unused software, faster, and more disengaged engineers. Paine's Smoot-Hawley history has no AI content. It is useful only as a reminder that trade restrictions on silicon can escalate in ways that no cost model captures.
What Changed
The compute supply curve looks contractually locked in, which is different from a forecast. SemiAnalysis cites slightly under $1T of binding, non-cancellable 15 to 20 year leases signed in 2026 year to date, with most of the debt already raised. More than 24GW across about 40 campuses is already in advanced construction for 2027 delivery. Prefabricated electrical modules, tent structures, and modular data halls cut build times below 10 months and MEP labor by up to 63%. Behind-the-meter generation bypasses grid interconnects that take 5 to 10 years. If this holds, the binding constraint moves from power and buildings to memory and logic in 2027 and 2028. That is a narrower, more geopolitically exposed choke point.
On the workforce side, nothing new was measured. What changed is that practitioners are now saying out loud that the job itself has changed, beyond productivity. Scheff's "they needed us" names the leverage that is eroding.
Cross-Expert Synthesis
Scheff and Jones describe the same transition from opposite emotional positions. Scheff mourns the loss of puzzle-solving as both pleasure and power base. Jones prescribes what replaces it: own who benefits, what they can do better afterward, and what finishing the whole job requires. Read together, they divide engineers into two groups. People whose identity rests on implementation lose status. People who already work past their role boundaries gain it. Neither offers data, so this is sentiment and heuristic, not proof.
SemiAnalysis supplies the economic driver both of them leave implicit. Falling inference prices and faster model cadence in 2027 and 2028 push artifact cost further down, which makes Jones's metric shift urgent rather than optional. The tension is that SemiAnalysis frames abundant compute as unambiguous progress, while the other two show that abundance of output is not abundance of value. More tokens running through an organization that measures output will raise waste, not results.
Enterprise Implications
- Inference pricing assumptions are stale. Contracts and business cases priced on 2025 and 2026 token economics will look expensive by mid 2027. Long-term committed spend at today's rates is the wrong bet.
- Output metrics become actively misleading. Agent-assisted teams can inflate PR counts and feature velocity without any matching gain in adoption. Boards will read that as success.
- Engineering adoption will split, not spread evenly. Expect outcome-oriented staff to adopt quickly and craft-oriented senior staff to resist or quietly disengage. Some of those senior people hold the most institutional knowledge.
- Technical authority moves. As implementation stops being the bottleneck, decisions on timelines and feasibility shift toward problem owners and verifiers. Org charts built around write, review, ship will misallocate authority.
- Hardware supply risk concentrates. A memory and logic shortage on top of export-control escalation makes accelerator supply a tail risk, not a line item.
What To Do About It
- Reprice before you commit. Keep inference and GPU commitments short (12 to 18 months) or index them to market rates. Push vendors for price step-down clauses tied to 2027 capacity.
- Change the scoreboard. For every AI-assisted delivery program, require Jones's three answers before funding: who it is for, what they will do better, and what full completion looks like. Report adoption and task completion alongside velocity, or instead of it.
- Redesign senior engineering roles explicitly. Route craft-oriented engineers toward verification, architecture, hard debugging, and agent output review. That keeps the puzzle work and puts their judgment where agents fail. Acknowledge the loss openly; telling people to "adapt" accelerates attrition.
- Change hiring rubrics now. Weight problem framing, review judgment, and systems ownership above raw implementation speed.
- Stress-test silicon exposure. Model at least one scenario where export controls escalate during the memory shortage, and identify which workloads could run on alternative hardware or smaller models.
Customer question worth asking: "If your agent rollout doubled code output next quarter, which metric on your dashboard would tell you whether anyone's job got easier?" Most cannot answer, and that gap is the engagement.
Risks and Blind Spots
SemiAnalysis sells the model it is defending, and its core data is proprietary. "Power is broadly solved" rests on behind-the-meter generation whose cost, emissions permitting, and reliability at gigawatt scale go unexamined. The oversupply question is raised and left unanswered. A 2028 glut is as plausible as a silicon shortage, and either one changes vendor stability.
Both workforce sources are short opinion pieces with no evidence. Jones's "work past your boundaries" ignores scope creep, burnout, and accountability gaps when everyone owns everything. Scheff's two-way split is probably too clean, since most engineers hold both motivations.
The Paine analogy is the weakest link. Using 1930s autarky to reason about chip export controls is scenario fuel, not prediction, and should not drive capital decisions alone.