Research Labs

AI research from the Costa Blanca.

Independent briefs on how artificial intelligence works and what it costs — token prices, compute supply, capital flows, the jobs it reshapes, European and Chinese regulation, cross-border dealmaking, and the engineering practice growing around AI tools. Every brief is sourced, dated, and free: read it in the browser or take the PDF.

The briefs

Published research.

Each brief opens with the numbers that matter, argues one thesis, and closes with a full numbered source list you can check yourself.

Pricing4 Oct 2026

Decision models stop charging for output and undercut LLMs on routine agent calls

Five firms shipped models that only make typed decisions, most with output billed at zero. They cut the cost of routine agent calls against frontier LLMs by two orders of magnitude, but barely beat small LLMs, and the benchmarks are mostly self-reported.

$0.042TypeSafe Jev price per million input tokens, with output free (15 Sep 2026)
Read the brief → ↓ PDF
Labour markets4 Oct 2026

One in five US job cuts now cites AI. The labour data has not caught up

AI is the leading stated reason for US job cuts in 2026, yet most of those announcements came before July and they are small against measured layoffs. The displacement signal sits in hiring and the labour share, not firing.

21%share of announced US job cuts attributed to AI, January–September 2026 (Challenger, 1 Oct 2026)
Read the brief → ↓ PDF
Infrastructure4 Oct 2026

Gigawatt delays push OpenAI and Anthropic toward 20 MW data centers

Anthropic and OpenAI are reportedly shopping for 20–30 MW sites because their gigawatt campuses will not be ready in time. Powered floor space, not GPUs, is now what rations AI supply.

20–30 MWsize of the data center deals Anthropic and OpenAI are reportedly exploring, against 460 MW to 8 GW leases signed earlier (CNBC, Sep 2026)
Read the brief → ↓ PDF
Market structure4 Oct 2026

AMD is buying a view of tomorrow's workloads

AMD will pay about $8.2 billion in stock for Fei-Fei Li's World Labs, a pre-revenue world-model lab valued at $5 billion in February. Chipmakers are now buying the labs that define their future workloads.

$8.2Ball-stock price AMD agreed for World Labs on 28 Sep 2026, 1.64× its Feb 2026 valuation
Read the brief → ↓ PDF
Pricing power4 Oct 2026

The frontier premium now lasts three weeks

Anthropic matched its own premium model at 40% of the per-token price three weeks after launching it. The premium tier's exclusive window halved in one release cycle, and frontier R&D now has to be paid back by volume.

21 daysfrom Claude Fable 5.1's launch to a Fable-level Opus 5.5 at 60% lower per-token prices (Sep 2026)
Read the brief → ↓ PDF
Energy and policy4 Oct 2026

Congress punts on who pays for AI's grid, so the states are writing the bill

The Senate stalled a voluntary federal standard on data center grid costs while California wrote a binding one. The largest bill so far, $29.4bn in PJM capacity costs, sits where neither approach reaches.

$29.4bncapacity costs attributed to data centers across PJM's last four auctions, per the market monitor (Jul 2026)
Read the brief → ↓ PDF
Policy & regulation4 Oct 2026

Antitrust law will not make Google pay for AI Overviews

A federal judge acknowledged that AI Overviews hurt publishers and still dismissed every antitrust claim against Google. US compensation for lost traffic now runs through Google's own pilot or through Congress, while the UK and EU take the regulatory route.

8% vs 15%Google searches ending in a click on a traditional result, with and without an AI summary (Pew, Jul 2025)
Read the brief → ↓ PDF
Policy and procurement4 Oct 2026

A usage policy is now a procurement liability

A 2–1 appeals panel held that Claude's built-in usage limits can make Anthropic a supply-chain risk under 41 U.S.C. § 4713. The ruling is narrow in reach, Pentagon work only, but broad in logic.

2–1D.C. Circuit vote denying Anthropic's petition against the Pentagon exclusion (25 Sep 2026)
Read the brief → ↓ PDF
Prices & labour4 Oct 2026

Voice AI's price war is being fought with expiry dates

Three voice-AI launches in nine days, three introductory prices that expire within three months. At list prices the speech layer still costs a fraction of an agent's hourly wage, which moves the automation bottleneck elsewhere.

~$1–3estimated speech-layer cost per conversation-hour at Oct 2026 prices, against a $21.53 median US agent wage
Read the brief → ↓ PDF
Open weights4 Oct 2026

Xiaomi gives away the second tier of AI

A phone maker released a trillion-parameter agent model with free weights and an API priced at a fraction of closed rivals. It matches the closed second tier, not the frontier, and that is the market it squeezes.

$0.13cost per Artificial Analysis index task for MiMo-V2.6-Pro, against $3.26 for GPT-6 Astra (Sep–Oct 2026)
Read the brief → ↓ PDF
Market structure4 Oct 2026

The frontier is now rationed, not priced

Three labs now charge the same price for frontier intelligence, and one of them will not sell its best model at any price. Argon shipped to an eligibility list of about 650 partners rather than a checkout page.

~650vetted partners with access to Gemini 4 Argon at launch, against an open API for everyone else (Oct 2026)
Read the brief → ↓ PDF
Policy & market structure4 Oct 2026

China stopped buying chips. It started renting them.

A reported $7bn, 100,000-chip lease moves no hardware across any border — and therefore trips no export control. The rental channel is now operating at roughly seven times the scale of the arrangement that made headlines nine months ago.

100,000chips reportedly leased, none of them exported (FT, 30 Sep 2026)
Read the brief → ↓ PDF
Competition law21 Sep 2026

Safety coordination just became an antitrust problem

In seven days the frontier labs asked permission to slow down together, the Senate refused it, and four subscribers sued them for doing it anyway. The binding constraint on release pacing is now US competition law.

4frontier labs named in the Sherman Act §1 class action filed 18 Sep 2026
Read the brief → ↓ PDF
AI economics14 Sep 2026

The frontier stopped getting cheaper

OpenAI and Anthropic released new flagships at exactly the same list price — $10 per million input tokens, $50 output — while the commodity tier collapsed to $0.15. Per-token price is now the wrong number to forecast.

$10 / $50identical list price per 1M tokens for GPT-6 Astra and Claude Fable 5.1 (Sep 2026)
Read the brief → ↓ PDF
AI economics5 Sep 2026

The token got cheaper. The task did not.

The usage-weighted price of a million LLM tokens halved between May and September 2026 while the frontier-tier list index rose 72% year on year. The number that matters is dollars per finished task, and it is still rising.

−52%usage-weighted token price, 31 May → 3 Sep 2026
Read the brief → ↓ PDF
Compute & supply27 Aug 2026

Memory, not GPUs, now sets the price of AI

Memory is now about 62% of a next-generation rack's bill of materials. Nvidia guided its gross margin to a 71–72% trough and warned its biggest customers that AI server prices rise more than 15% from early 2027.

62%memory share of a Vera Rubin rack's bill of materials, against 53% a generation ago (Aug 2026)
Read the brief → ↓ PDF
AI economics27 Aug 2026

Nvidia now underwrites the demand it forecasts

A $96.2bn quarter, the first year-ahead guide in company history — and $108.5bn of guarantees standing behind other people's data centres. AI's binding constraint moved from compute to credit.

$108.5Bmaximum guarantee exposure behind customers' data centre sites and leases (Aug 2026)
Read the brief → ↓ PDF
AI economics27 Aug 2026

When compute became a receivable

Nvidia booked 18% more revenue and collected less than half the cash. Receivables absorbed roughly $22bn in three months, supply commitments reached $279bn, and five customers now hold about 70% of the balance.

60 daysdays sales outstanding, up from 45 a quarter earlier (Q2 FY2027)
Read the brief → ↓ PDF
EU regulation24 Aug 2026

The AI Act's fine book just opened

Since 2 August 2026 the Commission can fine general-purpose AI providers up to 3% of global turnover. Nine of the ten leading labs signed the Code of Practice; Meta refused. The AI Office has about 165 staff to police it.

3%maximum fine as a share of global annual turnover, enforceable since 2 Aug 2026
Read the brief → ↓ PDF
Labour economics18 Aug 2026

The missing rung: AI's young-worker employment gap hits 19%

Stanford's revised payroll data widens the shortfall for 22–25-year-olds in AI-exposed occupations from 15% to 19% — while aggregate employment shows no displacement at all. The entire adjustment runs through hiring that never happens.

19%employment shortfall, ages 22–25, in the most AI-exposed occupations (Jun 2026)
Read the brief → ↓ PDF
Cross-border M&A18 Aug 2026

Beijing unwound a closed AI deal — and rewrote where a company comes from

China ordered Meta's completed $2bn Manus acquisition reversed eight months after closing, then wrote the reasoning into general law: jurisdiction now follows where a technology was built, not where the holding company is registered.

$2.0bnMeta–Manus deal closed Dec 2025, ordered unwound Apr 2026
Read the brief → ↓ PDF
AI economics18 Aug 2026

ChatGPT ad prices collapsed before OpenAI went global

The CPM fell from $60 to roughly $25 in nine weeks and the entry ticket from $250,000 to $10,000. The five-market expansion is a search for volume to offset a falling unit price, not a victory lap.

$60 → $25ChatGPT ad CPM, 9 Feb to 17 Apr 2026
Read the brief → ↓ PDF
Engineering practice12 Aug 2026

The spec is the new source code: how spec-driven development works

When AI agents write most of the code, the scarce skill becomes saying precisely what you want — as four versioned files a machine executes and a human reviews. New discipline, or waterfall in a new jacket?

124k+GitHub stars on Spec Kit, the open-source SDD reference toolkit (Aug 2026)
Read the brief → ↓ PDF
AI capabilities12 Aug 2026

Inside the loop: how recursive self-improvement (RSI) actually works

Every working RSI system is the same machine in different clothes: a generator that proposes, a verifier that scores, a selection rule that keeps the winners. Its power — and its ceiling — come entirely from the verifier.

20→50%SWE-bench score the Darwin Gödel Machine earned by rewriting its own code
Read the brief → ↓ PDF
Venture capital12 Aug 2026

The recursion premium: self-improving AI is repricing the startup world

AI now writes most of the code at the lab that builds it, and investors are pricing pre-revenue superintelligence bets above $25bn. Whether the loop closes is venture’s most expensive open question.

80%+of code merged at Anthropic now authored by Claude (May 2026)
Read the brief → ↓ PDF
AI economics12 Aug 2026

The price of intelligence fell 43% in ten weeks

Blended token prices collapsed in ten weeks — and almost none of it was a price cut. What buyers substituting into open-weight models means for $725–785bn of hyperscaler capex.

−43%blended token price, 31 May → 8 Aug 2026
Read the brief → ↓ PDF
Compute & supply12 Aug 2026

The memory tax: AI's constraint moved from GPUs to DRAM

Amazon raised 2026 capex by $20bn and named memory as the reason. AI’s binding constraint just moved to a market with three suppliers, multi-year lead times and no software workaround.

+$20BAmazon's 2026 capex raise, attributed to memory costs
Read the brief → ↓ PDF
EU regulation12 Aug 2026

The half of the AI Act that switched on

The EU switched on its AI labelling rules on 2 August 2026 and left the expensive half off until December 2027 — sixteen months of a rulebook Europe has chosen not to enforce.

16 monthsdelay of stand-alone high-risk duties, to 2 Dec 2027
Read the brief → ↓ PDF
AI economics6 Aug 2026

The subsidized intelligence economy

Frontier labs sell intelligence below cost, China sells it for a tenth of the price, and the largest capital buildout in American history is financed against demand that has not shown up.

−$14BOpenAI's projected 2026 loss, nearly 3× its 2025 loss
Read the brief → ↓ PDF

Method. Figures are attributed and dated inline. Where independent sources disagree, the disagreement is reported rather than averaged. Scenario probabilities are analytical judgement, not forecasts. Briefs are published by the Alicante Tech Vibes community and are free to read, quote and share with attribution.