Executive Summary
Today's three sources point to the same conclusion: in the next phase of AI capability, the advantage goes to whoever holds records that are tied to verified outcomes, not to whoever holds the most data or compute. Google paid $10M for a bankrupt airline's inbox. Periodic Labs is building robotic labs so it can generate data nobody can scrape. Chinese labs ship frontier models with no disclosed evaluation at all. The question underneath all three is who decides what "success" means. For enterprises, that decision is becoming the main point of leverage over vendors, and most are handing it away. Separately, Chinese open-weight releases will keep pushing prices down, and Chinese policy will not slow them. Unverified capability will keep getting cheaper, so buyers will have to do the verification themselves.
What Changed
- Work archives became a priced asset class before anyone validated them. Google outbid Mercor ($10M vs $7.5M) for Spirit Airlines' email and Teams corpus. The sale hearing is October 14. Brokers already list archives at $20K to $5M.
- Mercor is vertically integrating. It pairs data acquisition with the Deep Tune training-environment acquisition. Paid human experts now write the definitions of "task success" that later back vendor claims about which jobs agents can replace.
- The China safety gap is now quantified. Of 857 releases from China's nine leading labs, 94.9% carry zero safety disclosure, and no frontier text model has shipped with a dangerous-capability evaluation. China regulates at the application layer only; there is no compute or capability trigger. The state's deployment targets are 70% agent penetration by 2027 and 90% by 2030.
- "Pacing" is dead as a Western strategy. SemiAnalysis reports that Anthropic dropped its own pacing pledge within days, citing price pressure from Chinese open-weight models. That claim comes from one source and should be confirmed independently.
Cross-Expert Synthesis
Jones and Periodic Labs appear to disagree, and the disagreement resolves into the most useful idea in today's material. Jones argues that training on the full record of work teaches agents to imitate its performance: status updates, follow-up pings, meeting overhead. Periodic argues the opposite for science. Training only on final published results teaches models to "fake work" by recalling answers they already know, so the full process lineage, including negative results, is the moat.
The two positions fit together once you look at what each record contains. Periodic's lineage is instrumented: every step in the campaign connects to a physical measurement (XRD, SEM) that settles what actually happened. Spirit's archive has no such connection. A thread marked "resolved" may hide a failure discovered three weeks later, and the company that produced it went bankrupt. Process data is valuable only when it can be joined to ground truth. A full record without outcome labels is noise. A full record with outcome labels is a moat. Volume is not the variable that matters.
SemiAnalysis supplies the same idea at the model layer. Chinese weights are, in its words, "capability-tested by the market, not by any disclosed evaluation." Whether the problem is training data or deployed models, the missing piece is the same: nobody upstream owns verification. The work falls to the party that can observe outcomes, which is usually the enterprise.
The two training-data sources also agree on a quieter point: data is gaining value relative to compute. Periodic says this explicitly. Google's bid says it with money.
Enterprise Implications
An enterprise's operational exhaust (tickets, email, transcripts) is worth very little to anyone, including the enterprise, until it is linked to outcomes measured after the fact: refunds avoided, defects caught, renewals saved. Enterprises with mature systems of record that connect actions to results own something Google is currently paying distressed-asset prices to approximate.
Expect more vendor capability claims built on expert-defined success criteria that the customer never sees. Mercor-style ground truth will shape replacement narratives in procurement. Buyers who cannot inspect the rubric are buying an assertion, not a measurement.
Inference costs will keep falling on the back of Chinese open weights, and evaluation costs will become the buyer's line item. Any architecture that uses those models needs an internal evaluation and red-team layer. Nobody else will provide one.
Periodic's go-to-market (engineers embedded on site, fine-tuning on customer data, moving toward outcome-based pricing) is the template for vertical AI in physical and industrial domains, and it applies directly to advisory work.
What To Do About It
- Audit for outcome linkage before investing in any agent-training or fine-tuning program. Inventory which workflows can join records to results measured 30 to 90 days later. Fund only those.
- Prohibit the sale or licensing of internal work archives until target tasks, success criteria, and PII scope are defined. Unvalidated archives lose value, and a resale that underperforms damages credibility.
- Require vendors to disclose their success rubric. Ask who defined task success, against which outcomes, and with which negative cases. Treat a refusal as a red flag in diligence.
- Build an internal evaluation harness before standardizing on any open-weight model, Chinese or otherwise. Include domain-specific dangerous-capability and data-exfiltration probes.
- Position agents as augmentation, and say so to staff and unions. Jones's figure (up to a third of teams quietly resisting rollouts framed as replacement) is anecdotal, but the Spirit union's challenge shows it can become a legal problem.
Customer question worth asking: "Which of your records can you tie to an outcome you measured weeks later? That subset is your only proprietary AI asset. Everything else is exhaust."
Risks and Blind Spots
Jones's thesis may be wrong at Google's scale. With enough volume, models may learn to tell the judgment moments apart from the coordination noise without explicit labels, and the $10M may prove cheap. The adverse-selection argument also cuts both ways: failure records with known outcomes are themselves labeled negatives, which Periodic says it values.
SemiAnalysis measures disclosure, not capability or safety. Undisclosed is not the same as unevaluated, and the data cannot separate the two. The claim about Anthropic's pacing pledge is load-bearing for the strategic conclusion and needs independent confirmation.
Periodic's lessons come from materials science, where characterization tools produce clean ground truth. Most knowledge work has no XRD equivalent. Constructing outcome labels for business processes is a large, mostly unpriced cost.
The October 14 Spirit hearing will set early precedent on worker standing over training-data scope. Watch whether the court narrows PII scope or recognizes the union's objections to what models can infer from the data. Either outcome would change how much internal archives are worth across the market.