Executive Summary
Today's sources agree on one shift: producing software and agent output is no longer the scarce input. Three data points show it:
- Berman built an Excel clone for about $10 in tokens.
- Warp's cost per PR fell from $80 to $15–20.
- A 27B model on a laptop reproduced a paper result credited to frontier models.
What is scarce now is three things. Can you show the agent's output is correct? Who maintains it? Can a user trust it without watching it? On supply, SpaceX/xAI is becoming a fourth hyperscale-grade compute source. Its build cycles run about 91 days against an 8–12 month industry norm. That gives buyers leverage on price and timing, but it brings a new vendor-conflict risk. Enterprises should make the model, the harness (the agent runtime) and the compute swappable, and should invest in the evaluation loop rather than in any single model.
What Changed
- Capability threshold. Berman ties the flip from buying to building software to Opus 5.5. He describes it as a step change, but his evidence is a single test.
- Cost-tuning headroom has a number. Warp cut cost per PR 4–5x by taking Opus out of triage and removing redundant sub-agents. Which model handles which step is now a cost decision.
- New compute supply. SpaceX targets ~13GW by 2027, with 3.6GW under construction. It runs on its own on-site gas generation plus a ~6GWh battery system, and TeraFab may offer capacity to third parties. These are SemiAnalysis estimates, not company disclosures.
- Frontier-only claims are being checked locally. Willison reproduced the no-tools weight-reading result on Qwen 3.8 27B.
Cross-Expert Synthesis
Berman and Lloyd agree that generating software is cheap. They disagree on where the value goes. Berman says software value approaches zero and always-on agents will absorb maintenance. Lloyd's business assumes the opposite: agent output is worth little until it is traced, scored and benchmarked against your own code. For enterprises, Lloyd has the stronger case. Berman never prices operations, security or support. Lloyd's example (95% correct at $6 per task versus 93% at $2) is the kind of evidence that settles which model to use. Berman's thesis holds for commodity single-user tools, not for systems of record.
Jones adds the user side of the same loop. Muse won on polish with a model that is not the strongest. Meta tuned 100+ behaviors and made the agent's status visible. That is Lloyd's automated quality scoring applied to user experience. Both reject benchmark rank as the deciding variable. The real tax is the cost of supervising the agent: human touches per PR for developers, and "is it working on my task?" for users. Jones also qualifies Berman. Cheap first drafts tempt teams to keep starting new things and never finish one. A sprawl of vibe-coded internal tools is that same failure across a whole company.
Willison supplies the method. Separate what a model can do on its own from what it can do with tools. Reproduce claims on your own hardware. Do not trust a model's stated confidence. Apply the same discipline to Warp's self-reported metrics and to SemiAnalysis's 300B+ ARR projection.
SpaceX connects back to Lloyd's swappable inference layer. A new neocloud with a different cost base only helps if workloads can move to it. SpaceX also owns a frontier lab (xAI) while selling capacity to Anthropic and Google, so the ability to move workloads becomes a governance requirement, not just a cost option.
Enterprise Implications
Leverage at renewal on commodity per-seat SaaS is now real. Expect vendors to defend their pricing with data, integrations and compliance.
For platform teams in 10–100-engineer orgs, Lloyd's "software factory" (the whole agent pipeline defined as version-controlled config) is a usable reference design whether or not they buy Warp. Large shops with their own platforms will build it themselves.
Compute buyers now have a credible option outside AWS, Azure and Google. SpaceX diligence must cover:
- Air permits. Southaven's turbine cap drops from 2.7GW to 1.24GW by July 2027.
- The ~300MW power gap. It depends on batteries cycling against gas.
- Separate service tiers for SpaceX's own training versus customer clusters.
- Data isolation from a vendor that also runs a competing lab.
Agent programs will succeed or fail on how visible and well-tested the agent's behavior is, not on model choice.
What To Do About It
- Measure before scaling. Set up a central store of agent traces and a way to replay real runs before expanding agent use. Benchmark model and harness combinations on the client's own repos, then route each task to the best cost-versus-correctness option.
- Keep every layer swappable. Make inference provider, sandbox and harness pluggable, including any harness that speaks the ACP standard and bring-your-own provider. This burns down existing vendor commitments and keeps neocloud options open.
- Build a SaaS renewal kill-list. List commodity tools where an internal build plus agent maintenance costs less than the license. Before any replacement ships, require a named owner, a security review and a commitment on patch turnaround.
- Add SpaceX/xAI to the compute RFP shortlist, with a diligence addendum on conflict of interest and permit exposure.
- Talk track: "If you can't tell me your cost per merged PR by SDLC stage, you're not running an agent program. You're running a subscription."
Risks and Blind Spots
Nearly every number today is unaudited. Warp's metrics are self-reported, SemiAnalysis works from satellite imagery and permit filings, and Berman's benchmark is a single run.
Many teams' current economics rest on flat-rate coding subscriptions, which Lloyd puts at 5–50x cheaper than API billing. That gap could close at the next terms-of-service enforcement, which would change the math on internal builds.
Software cloned by agents raises IP and decompilation exposure that legal teams have not assessed.
Weight-reading has only been shown on toy networks. Treat any vendor pitch built on it as unsupported.
SpaceX's on-site gas generation makes its capacity more exposed to emissions regulation than grid-connected capacity.