Newsletter

Four AI releases landed in six days, and not one of them competed on raw model capability. Anthropic’s Claude Duet runs on Opus 5.5, a model it had already shipped, wrapped in 6,000 connectors and a daily briefing. Google open-sourced a framework that improves results by rewriting the prompts, tools and memory around a model whose weights never change. Reflection gave away 501 billion parameters under Apache-2.0. DeepSeek published 763 billion under MIT, then put a meter on the search API sitting next to it.

The weights are becoming the giveaway and the scaffolding is the product. That is a procurement problem before it is a technical one. If the model is free or close to it, what you are buying is connectors, approval gates, audit trails and the people who can wire them together. None of that comes with a benchmark score, and all of it costs more than tokens.

Two of this week’s four sets of numbers come from vendors measuring themselves, which is worth holding lightly. What would complicate the picture is a lab shipping a clearly better agentic model and keeping it closed. Beam’s 81.6% on SWE-bench Verified suggests nobody currently believes that moat would hold.

---

Anthropic unveils Claude Duet AI agent app

Reuters · 21 October 2026 — Duet watches connected apps and raises what needs attention rather than waiting for a prompt. It is a waitlisted research preview on iOS, Android and claude.ai for US Pro and Max subscribers, built on Opus 5.5 with over 6,000 connectors including Gmail and Slack.

Why it matters: It takes the always-on agent pattern covered here on 1 October from OpenAI’s cloud-hosted Dots to a mainstream subscription tier — and Anthropic concedes the agent can misjudge urgency.

Beam is now available

Reflection · 16 October 2026 — Reflection put Beam’s weights on Hugging Face under an Apache-2.0 licence: 501.3 billion total parameters, 23.4 billion active, 1M context. Its own model card reports 81.6% on SWE-bench Verified and 80.1% on Terminal Bench v2.1.

Why it matters: A permissive licence on a model scoring in frontier territory at agentic coding leaves very little room to sell weights on their own.

DeepSeek-V4.1-Flash-1017 ships, and Search and Fetch open to everyone

DeepSeek · 17 October 2026 — 763 billion parameters under an MIT licence, 1M context, output up to 384K, at $0.15 and $0.60 per million tokens off-peak. The same post opened the Search and Fetch APIs at $0.75 per 1,000 calls, with billing from 31 October.

Why it matters: Free weights, metered retrieval plumbing. The whole shift sits inside one announcement.

Google Open-Sources Harness2, a Recursive Agent Harnessing Framework

MarkTechPost · 18 October 2026 — Harness2 edits eight things around a frozen model: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. The Apache-2.0 release reports GLM-5 under OpenCode rising from 64.2 to 75.7 on WorkBuddy-Bench — figures from Google’s own paper, not independent testing.

Why it matters: If editing the wrapper buys eleven points on a benchmark, the wrapper is where the engineering effort goes.

---

If you are budgeting for AI next quarter, budget for integration rather than tokens.

Written by my AI assistant.

Three US regulatory moves in the first week of October share a cause none of them names. PJM is 6.8GW short of the capacity it needs for the 2028-29 delivery year, and the bill for closing that gap has to land on someone.

The Department of Energy told FERC on 7 October that PJM’s reliability backstop would shift large-load costs onto existing ratepayers. Maryland’s regulator awarded 440MW against an 800MW storage target, into a market where PJM capacity has been clearing at its $325/MW-day cap. MISO, a separate market, asked FERC to fast-track studies for loads above 200MW — but only if they arrive paired with their own new generation.

That last condition is the tell. Regulators have stopped treating large loads as a connection-queue problem and started treating them as a cost-allocation problem. It is the same question Michigan settled bilaterally last week, when it made Google commit to 80% minimum billing for 20 years. Bilateral deals do not scale. Tariffs do.

On a separate axis, China’s commerce ministry proposed export controls on 12 solar manufacturing technologies. Note what is not covered: finished cells and modules. The target is process know-how and the equipment that embodies it, at the moment the US and India are building domestic factories.

Two things would confirm the reading: whether PJM’s 29 October filing allocates backstop costs to large loads directly, and whether China’s proposal clears consultation on 15 November with its equipment thresholds intact.

---

China proposes export controls on 12 solar manufacturing technologies

pv magazine · 16 October 2026 — The commerce and science ministries would ban exporting perovskite-silicon tandem processes at 30% efficiency or above and sub-70-micrometre wafer technology, and put TOPCon, HJT and back-contact steps under licence. Consultation closes 15 November; a registration rule took effect 16 October.

Why it matters: Finished modules are untouched. What is being fenced in is the ability to build a solar factory outside China.

DOE presses PJM on who pays for large loads

Utility Dive · 8 October 2026 — DOE filed a statement of position at FERC on 7 October arguing PJM’s backstop procurement would unfairly shift large-load costs to existing ratepayers — apparently its first FERC filing in at least five years. A revised PJM plan is expected by 29 October.

Why it matters: A federal department litigating cost allocation moves the venue for every data-centre interconnection in PJM.

Maryland’s first storage auction lands 440MW of an 800MW target

Utility Dive · 6 October 2026 — The Public Service Commission awarded Flatiron’s 400MW/1,600MWh Chalk Point project and REV Renewables’ 40MW/160MWh Jade Meadow III on flat 15-year terms, together adding 6.98 cents a month to the average residential bill. It passed over a 500MW/3,000MWh bid.

Why it matters: The state reached 55% of target while declining a bid that would have cleared it. Why it declined is not in the public record.

MISO asks FERC for a 120-day fast track for 200MW-plus loads

Utility Dive · 5 October 2026 — Loads above 200MW could leave the standard queue if paired with new generation in the same local resource zone, with both committed for 15 years and in service within three. MISO wants up to 10 applications per study period, three periods a year.

Why it matters: Bring your own generation is becoming the price of fast interconnection.

---

One story was dropped on age: BloombergNEF’s first-half investment figures, widely recirculated in late September, were published on 26 August and fall outside this brief’s 30-day window.

Written by my AI assistant.

The sector’s answer to its labour shortage is mobility: bring people in from oil and gas, mining and construction, and convert them. OPITO’s first Asia-Pacific survey says that answer keeps hitting a wall that has nothing to do with skill. More than two-thirds of employers there demand extra training, certification or assessment before they will take a worker from an adjacent safety-critical sector. Among workers who say they won’t move, 59% give the same reason: they doubt their existing training would be recognised.

Two weeks ago this brief covered a six-week conversion course offered as wind’s fix for its technician gap. A conversion only works if the certificate at the end of it is accepted somewhere else. The paperwork is moving the other way. Britain’s Electrotechnical Assessment Specification added four low-carbon work categories on 1 October, each requiring a formal Level 3 qualification and CPD evidence. Defensible on its own terms, and one more certificate that stops at a border.

It is also worth noting where competence actually failed this month. The most expensive British safety judgment of the fortnight, £4m, was not about a technician’s certificate. It was a routine lifting job that no competent person had planned.

What would complicate this reading: a recognition scheme people actually use. The UK’s Energy Skills Passport had fewer than 500 applications by December 2025. If OPITO’s mapping work lands differently, it will show up as employers dropping re-assessment, not as another framework.

---

Over two-thirds of APAC employers retest workers from adjacent sectors

Global Mining Review · 14 October 2026 — OPITO surveyed 390 workers and employers across Australia, Indonesia, Malaysia, the Philippines, Thailand and Vietnam. 57% said trust breaks down when training or certification is not recognised. CEO Stephen Marcos Jones wants “a shared language of safety, so skills can travel more widely”.

Why it matters: Every workforce plan in the sector assumes workers can cross from one industry to another; this is the first hard measure of what that crossing costs.

Four low-carbon work categories became competence-assessed on 1 October

NAPIT · effective 1 October 2026 — The updated Electrotechnical Assessment Specification adds EV charging, solar PV, electrical energy storage and micro wind turbines as separate categories, each needing a Level 3 qualification and CPD records. Periodic inspection and testing now requires two years’ practical experience.

Why it matters: Competence for renewables work in Britain is now assessed per technology rather than assumed from general electrical registration.

A £4m fine for a lift nobody had planned

HSE · 16 October 2026 — City Electrical Factors was fined £4m plus £6,304 costs after a 231kg switchboard load fell from a forklift onto a 40-year-old worker acting as a spotter. He now needs 24-hour care. There was no risk assessment and no lift plan. Inspector Keeley Eves called it “completely avoidable and caused by a failure to properly plan”.

Why it matters: The failure was in planning and supervision, not in the certificates held by the people on the floor — which is where most competence spending goes.

Canada put C$100m into Nova Scotia offshore wind, none of it into people

Global News · 15 October 2026 — Ottawa committed C$100m to the Wind West project, including about C$5m for wind and ocean measurement and nearly C$3m for transmission modelling and marine surveys. Nova Scotia wants to license 5GW by 2030. Minister Tim Hodgson said “nearly 30 countries are ahead of Canada”.

Why it matters: The named allocations are all data and survey work. A province starting from almost no offshore wind workforce has not yet costed the one it will need.

---

The through-line is unglamorous. Capital and capacity targets scale quickly; the machinery that lets a competent person work in a second country or a second technology does not. Watch whether recognition schemes start retiring assessments rather than adding them.

Written by my AI assistant.

Two unrelated companies spent this week attacking the same number, and it was not intelligence. Agility Robotics cut Digit 5’s charge time to nine minutes against 90 minutes of runtime, moving its run-to-charge ratio from 2:1 to 10:1. Hyundai, separately, said its logistics robots now charge in about 12 minutes instead of roughly two hours. Neither announcement makes a robot smarter. Both make one cheaper to own.

That convergence is the story, because duty cycle is what a buyer actually pays for. The humanoid sector has spent the autumn quietly retreating from general capability toward whatever can be measured: wheels in late September, a four-fingered hand with a production target two weeks ago, charge rate now. Each step is a narrower, more honest engineering claim than the one before it.

The awkward part is who is not in a hurry. Hyundai owns Boston Dynamics, and its own roadmap puts humanoids in new factories only after 2032, with vehicle wiring and interior trim as the target tasks. The energy problem is being solved faster than the purchasing decision. What would confirm the shift: whether Agility’s early-access customers in the first half of 2027 are named, and whether anyone publishes measured uptime rather than a claimed 20 hours in 24.

---

Agility picks Nyobolt for a nine-minute Digit 5 charge

The Robot Report · 15 October 2026 — Agility Robotics selected Cambridge battery firm Nyobolt for Digit 5: nine minutes of charge for 90 minutes of work, a 10:1 run-to-charge ratio against 2:1 on Digit 4. Early access is expected in the first half of 2027.

Why it matters: Chief business officer Daniel Diez said “Digit’s uptime is the business case” — charge rate, not dexterity, is the spec being priced.

Hyundai puts factory humanoids after 2032

Aju Press · 18 October 2026 — Hyundai Motor Group will introduce humanoids and full-scale automation only at factories built after 2032, with autonomous systems at its Singapore innovation centre by 2030. Its logistics robots already charge in 12 minutes instead of about two hours, at roughly 30% lower production cost.

Why it matters: The largest industrial buyer in this conversation has put humanoids seven years out, so supply is improving faster than demand.

Greater Bay commissions a GWh-scale solid-state line

Gasgoo · 15 October 2026 — Greater Bay Technology commissioned what it calls the industry’s first gigawatt-hour-scale all-solid-state line at Nansha on 13 October, retrofitted from part of an existing 4GWh liquid-lithium line. Cells are rated 260-480 Wh/kg and passed nail penetration without fire.

Why it matters: Converting an existing line rather than building one is how solid-state stops being a pilot programme.

Industrial Next retools lines in minutes, not weeks

TechCrunch · 13 October 2026 — Industrial Next, founded in 2021 by two former Tesla engineers, says its self-contained robot stations are taught in simulation rather than scripted, so lines retool in minutes. Its film-peeling station runs on a leading smartphone; BMW’s technology office is evaluating the system.

Why it matters: Changeover time, not cycle time, is what keeps delicate assembly work in human hands.

Applied Materials opens a $5bn shared R&D clean room

Fortune · 13 October 2026 — Applied Materials unveiled EPIC, a $5bn, 600,000-square-foot R&D facility in Sunnyvale, with Intel, TSMC, Micron, SK Hynix and Samsung placing research teams on site. The company says shared early access could cut concept-to-commercialisation from 10-12 years to six.

Why it matters: Direct rivals sharing one clean room is a bet that tool development, not fab capacity, is the binding constraint.

---

Nothing above is about intelligence. All of it is about how long a machine can stay on the floor, and how fast the floor can be changed.

Written by my AI assistant.

ANALYSIS — Michigan’s regulators conditionally approved DTE Energy’s contracts to power Google’s 1GW data centre in Van Buren Township on 1 October. The terms: a 20-year primary supply agreement, minimum monthly billing set at 80% of contract capacity, and a termination payment covering at least 15 years of those charges. Service is not expected to begin until December 2027.

That is the same instrument France reached for earlier this week, and it is starting to look like the only one that clears. France is auctioning winter 2030-31 capacity before it sells 2027’s. PJM, the biggest US grid, suspended the 6.8GW data-centre auction it had built around 15-year contracts. What has changed since Wednesday is who carries the risk, not whether it gets priced: France is socialising it through a regulated auction, Michigan has pushed it onto the buyer, and PJM has not answered the question at all.

Compute is pricing the same way. Gemini 4 Argon is cheap until 1 December, and the memory that AI servers run on is still getting dearer into the fourth quarter. Supply that takes years to build is meeting demand that arrives in months. In both markets the discount on the present is a promotional window, not a price.

The thing to watch is PJM’s next filing. If the long-dated contract is the only product on sale, the fight moves to who is able to sign one.

---

Michigan approves 20-year DTE contract for Google’s 1GW data centre

Planet Detroit · 1 October 2026 — The Michigan Public Service Commission unanimously granted conditional approval, Michigan’s first hyperscale contract decided through a contested case. Google also pays for up to 1,600MW of renewables and 480MW of DTE battery storage. Attorney General Dana Nessel had pushed for 90% minimum billing, citing stranded-asset risk.

Why it matters: Commissioner Katherine Peretick asked before the vote “How do we make sure that existing customers are protected?” — the question every regulator now has to answer in contract terms rather than tariffs.

PJM shelves its 6.8GW data-centre auction after FERC pushback

Bloomberg, via Energy Connects · 30 September 2026 — FERC called the plan “deeply flawed” and asked for a new proposal. The one-time auction would have used 15-year contracts to cover a 6.8GW shortfall. Data centre demand has added close to $30bn to recent grid-wide auctions. No new timeline has been set.

Why it matters: The regulated route to long-term capacity has stalled in the market where the demand is largest, which is why bilateral deals like Michigan’s are setting the terms instead.

A gas-equipment firm books over 1GW of power that bypasses the grid

Investing.com · 13 October 2026 — Enerflex won a contract for more than 1GW of behind-the-meter gas generation at a North American data centre, taking its behind-the-meter backlog past 1.75GW with deliveries running through 2028. The customer was not named. It authorised about $100m of facility investment.

Why it matters: Buyers who cannot get a grid connection on their timeline are building their own supply, which removes them from the capacity auctions that were meant to price them.

Memory suppliers are telling buyers the shortage lasts past 2029

TechRadar · 6 October 2026 — TrendForce says DRAM supply is very tight and fourth-quarter contract talks point to meaningful increases. A YMTC employee told The Wire China the shortage will last at least another three years. German DDR5 is reported up 544% on pre-crisis prices.

Why it matters: This is the input cost behind AI pricing that rises on a date rather than falling with scale, and it is contracted quarter by quarter.

---

Four stories, all of them about the same thing: nobody is selling the near term any more. If you buy power or compute on anything other than a multi-year commitment, you are now the marginal buyer.

Written by my AI assistant.

Google made Gemini 4 Argon broadly available on Tuesday at $2 per million input tokens and $10 per million output. Those are introductory rates. On 1 December they become $4 and $20. The model’s headline use is the one that consumes tokens fastest: persistent tasks that keep running while you are away from the machine, wired to more than 40 external apps.

Four of today’s five stories are about the meter or the gate, not the capability. That is not coincidence. Long-running agents are the first AI product where both the cost and the liability scale with how long the thing runs rather than with how often you ask it something. So Google prices runtime and puts a deadline on the discount. Workday makes the permission envelope itself the product. JPMorgan declines to issue a key at all. Liquid AI’s answer is to move the work off the meter and onto a laptop.

The Purdue paper is the uncomfortable one. Three repetitive habits turn up in 79 to 98 percent of coding tasks and account for up to 22.75 percent of what the task costs. At introductory prices that is a rounding error. At December prices, on an overnight run, it is a budget line somebody has to explain.

So watch 1 December. If agentic volume on Argon holds through a doubling, the overnight-agent pitch is a product. If it drains to on-device models or to whoever is still discounting, it was a promotion.

---

Google makes Gemini 4 Argon broadly available at $2/$10 per million tokens

VentureBeat · 13 October 2026 — Argon reached the paid API, AI Studio and Antigravity at $2 input and $10 output per million, cached input at $0.10. The rates expire 30 November and double from 1 December. It leads the Vals Index at 68.90% but scores 53 on Artificial Analysis, under Claude Opus 5.5 at 58.

Why it matters: The cheapest route to frontier-grade agent runs has a six-week shelf life.

JPMorgan blocks outside AI agents from directly accessing accounts

CNBC · 13 October 2026 — Jamie Dimon told the bank’s third-quarter earnings call that JPMorgan will not let external AI agents directly access or operate retail customer accounts, citing the risk of rogue agents. CNBC names Muse and Instinct among the tools affected.

Why it matters: The largest US bank has settled that agent access is a liability question before it is a product question.

Workday opens HR and finance data to agents with nobody logged in

Workday · 13 October 2026 — The Trusted Agent Platform now exposes Workday data to any MCP-compatible agent headlessly, with Agent Harness generally available. Every request carries Workday’s own permissions, approvals and audit trail. Headless can’t mean lawless, said CTO Gabe Monroy. Workday counts 11,500 customer organisations.

Why it matters: Two weeks after an agent control plane shipped free and open-source, the commercial version arrives with the permission model as its selling point.

Liquid AI ships a 24B agent model that activates 2B parameters

Liquid AI · 13 October 2026 — LFM2.5-24B-A2B activates 2B of 24B parameters per token, holds 131K context and runs 184 tokens per second on an Apple M5 Max. Vendor benchmarks put it at 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B, while losing on SWE-Bench Verified.

Why it matters: If agent runtime is what gets metered, the cheapest runtime is on hardware you already own.

Analyzing and Mitigating Cost-Inefficient Behaviors in Coding Agents

arXiv (Purdue University) · 25 September 2026 — Researchers traced 1,200 agent trajectories across 300 SWE-bench Verified tasks and found three repeated behaviours — subsumed retrieval, similar script generation and test re-execution — in 79 to 98 percent of tasks, accounting for up to 22.75 percent of task cost.

Why it matters: A fifth of the bill buys work the agent has already done once.

---

Nothing this week raised the ceiling on what an agent can do. Everything this week changed the terms on which you are allowed to let one run.

Written by my AI assistant.

Three governments moved in the same direction over the past fortnight, and not one of them did it by raising a price. France’s regulator set the parameters for a capacity auction four years ahead of delivery. Victoria bolted a fixed availability payment onto a 20-year contract for difference. Eskom stopped acting purely as a buyer of power and offered to put 2GW of its own solar and 1GW of its own batteries into a co-investment panel.

The common move is a shift from paying for energy to paying for being there. The reason is legible in French spot prices: roughly 800 hours at or below zero so far in 2026, against 359 negative hours across the whole of 2024. Where that is the shape of the market, a plant living on merchant revenue cannot be financed, and no amount of auction volume fixes that.

The 9 October edition noted 41.2GW of offshore wind under construction against just 2.1GW reaching final investment decision this year. What changed this week is the response on offer: not a better price, a different allocation of risk. Capital is still available for anything with a contracted floor — Grenergy’s $457m was non-recourse against a 12-year PPA with an investment-grade buyer.

Watch France’s cap. If the 2030-31 auction clears anywhere near €72/kW, the cost of firmness stops being a modelling assumption and becomes a published number other regulators have to argue with.

---

France defines terms for long-term capacity auction

pv magazine · 12 October 2026 — Regulator CRE proposed an 82GW target for France’s first long-term capacity auction, covering winter 2030-31, with 3GW eligible for multi-year contracts and price caps held at €72/kW and €15/kW. Grid operator RTE put additional capacity needs at 0 to 2.5GW.

Why it matters: Under the current timetable bidders must commit to 2030-31 before contracts exist for the three winters in between.

France hits record 800 hours of negative, zero electricity prices

pv magazine · 2 October 2026 — Storio Energy’s September monitor counted about 800 hours of negative or zero spot prices in France this year, an all-time high, against 359 negative hours in all of 2024. September’s average daily spread reached €214/MWh, the widest since 2022.

Why it matters: This is the revenue profile every capacity payment and availability payment is being designed around.

Australia’s first offshore wind auction takes shape with longer contracts and special payments

RenewEconomy · 6 October 2026 — Offshore Wind Energy Victoria set 20-year contracts for difference for the first 2GW in the Gippsland zone, against a 10-to-15-year global average, plus a fixed Availability Payment that starts only after commercial operation. The strike price is unannounced.

Why it matters: Contracts are due in 2028, construction in 2032 and power in 2034, so the terms are being set a decade before delivery.

Eskom kicks off tender for 6 GW of renewables-plus-storage

pv magazine · 12 October 2026 — Eskom opened a request for qualification for up to 6GW of utility-scale renewables and storage, committing at least 2GW of solar and 1GW/4GWh of batteries of its own. Successful developers join a five-year co-development panel, each project ring-fenced in its own SPV.

Why it matters: A state utility under financial strain is choosing to take equity risk rather than only sign power purchase agreements.

Grenergy secures $457 million financing for 446 MW of solar and 1.8 GWh of storage in Chile

Renewable Energy Magazine · 9 October 2026 — The Spanish developer closed $457m of non-recourse debt across 446MW of solar and 1.8GWh of storage. The anchor is Algarrobal, 242MW and 1.5GWh, on a 12-year hybrid PPA with an investment-grade offtaker from January 2028. CEO David Ruiz de Andrés called it “the largest financing transaction in our company’s history”.

Why it matters: Lenders priced the contracted hybrid, not the merchant upside.

---

The instruments differ by jurisdiction, but the direction is one way: the public balance sheet is becoming the counterparty of last resort for firm capacity. The number worth waiting for is a cleared capacity price.

Written by my AI assistant.

The UK Health and Safety Executive is investigating one death in Great Britain and knows of a second outside it: workers who fell through non-load-bearing glass floor windows after lifting off the gratings that cover them to clean the glass. HSE’s bulletin notes that industry design standards require those glass floors to be covered, but do not explicitly require the cover to be secured. G+, the offshore wind sector’s safety body, circulated it to members on 8 October.

Three of today’s items describe the same failure mode, and it is not a failure of rule-writing. The sector’s safety assurance is documentary, and little of it is verified in the field. The design standard specified a cover and stopped there. G+ published Lifesaving Rules for offshore wind two years ago and is now surveying members to find out whether they are actually used. Researchers scanning for exposed renewables controls found turbine pages with live Start, Stop and Reset buttons, and report that most operators cannot say who holds remote command of their own assets. A written control nobody checks is a control that exists only on paper.

What would settle it is cheap to observe. HSE asks for three things: gratings fixed so removal needs tools, a risk assessment covering cleaning, and signage at every non-load-bearing pane. Whether those turn up in the next inspection rounds, or only in the next set of documents, is the test.

---

HSE: potential for falls through non-load-bearing glass floor windows

Health and Safety Executive · September 2026, circulated 8 October — Bulletin ED03-2026 covers crane and drilling derrick cabins where glass floors improve the operator’s view. Two fatal falls followed workers removing the gratings to clean the glass. Routine inspections keep finding unsecured gratings.

Why it matters: Offshore wind inherits this equipment and these cabins from oil and gas, including the gap in the standard.

8,547 wind and solar systems found facing the open internet

Help Net Security · 9 October 2026 — Modat and the Dutch cyber centre NCSC-NL tied 8,547 internet-facing systems to specific wind and solar sites across 35 countries: 7,942 solar, 605 wind. One page exposed working Start, Stop and Reset controls.

Why it matters: A remote stop command reaching a turbine while someone is working in it is a safety problem before it is a security one. Tributech’s Thomas Plank says of a full inventory of remote connections: “Many operators can’t produce that today.”

G+ asks its own members whether the Lifesaving Rules are being used

G+ Global Offshore Wind Health and Safety Organisation · 24 September 2026 — Two years after publishing a set of Lifesaving Rules written for the offshore wind risk profile, G+ has opened a survey to find out how they are being used across its membership.

Why it matters: Asking is the right move, and the fact that it has to be asked is the finding.

Flotation Energy cuts 15-20 UK roles as Green Volt leaves consenting

reNEWS · 12 October 2026 — The TEPCO-owned floating wind developer is making 15 to 20 of roughly 130 UK staff redundant as its 560MW Green Volt project moves from development and consenting into detailed design and procurement.

Why it matters: Development-stage headcount is the first thing cut when a project changes phase, and those are the people who carry the site knowledge forward.

---

The common thread is cheap to state and expensive to fix: the sector can write a standard faster than it can confirm anyone is following it. Secured gratings are a bolt. Knowing who can stop your turbine is a list. Neither is hard. Both are currently assumed.

Written by my AI assistant.

Offshore wind’s problem this quarter is not construction. TGS | 4C counts 41.2GW being built right now and more than 12GW entering service this year — the yield of decisions taken in 2021 and 2022. What has stopped is new commitment. Just 2.1GW has reached final investment decision in 2026, and even the year-end forecast of 5.6GW would leave this the second-lowest year since 2020.

Three of today’s stories point at the same shift from different directions: the state has become the marginal investor. Scotland is putting a £1.45m grant behind £5m of private money and eighty jobs at Montrose — real, and small enough to show how incremental the £500m fund covered here on Wednesday actually is. Washington has reimbursed $3.9bn to six developers to take back 19.2GW of leases, with some of that capital heading to LNG instead. Public money is opening a factory on one side of the Atlantic and closing a pipeline on the other. Private capital is doing neither at scale.

What would settle this is the next two quarters of auctions — AR8 in the UK, and the German and Dutch rounds. If FID volumes stay near 5GW a year while 40GW finishes building, the 2030 targets stop being a supply chain question and become an arithmetic one.

---

Offshore wind’s 2026 pipeline reveals a widening gap between construction and new investment

TGS · 21 September 2026 — TGS | 4C’s quarterly Global Market Overview counts 41.2GW of offshore wind under construction worldwide against 2.1GW that has reached final investment decision in 2026, with 5.6GW forecast by year-end. Steel costs 53% more than a year ago.

Why it matters: Construction volume reflects decisions already taken. FID is what tells you about 2030, and it is near a six-year low.

New critique of offshore wind clawbacks: taxpayers lose out

Roll Call · 30 September 2026 — Taxpayers for Common Sense says Interior has reimbursed $3.9bn to six developers to buy out offshore wind leases covering 19.2GW of nameplate capacity, including $1.22bn to RWE’s US offshore arm and $928m to TotalEnergies. Both are directing capital to LNG projects.

Why it matters: The buyouts turn collected public revenue into cancelled capacity — the same state-as-marginal-investor mechanic as Montrose, running in reverse.

Subsea Micropiles to invest £5m in Montrose operations

Aberdeen Business News · 28 September 2026 — Subsea Micropiles will spend £5m over three years expanding at Montrose Port, backed by a £1.45m Scottish Enterprise grant, to build anchor and foundation systems. Around 80 jobs, phased over three to five years. Chief executive Derek Robertson: “Scotland has become an important base for our growth.”

Why it matters: This is what Wednesday’s £500m fund looks like at ground level — eighty jobs at a time, over five years.

Cheaper AI models, bigger AI bills

Fortune · 23 September 2026 — Reporting on McKinsey’s State of AI in 2026 survey, Fortune notes that the same agent task can cost up to 30 times more from one run to another, even as capability-adjusted model prices fall.

Why it matters: Thursday’s 90% price cut sets the floor, not the bill. Variance of that size is a budgeting problem, not a pricing one.

---

Both halves of this week share a shape: a headline number that is real, and a committed number that is much smaller. Worth watching which one the 2030 plans were built on.

Written by my AI assistant.

Three of this week’s four announcements were about everything except the model’s reasoning. Anthropic cut Haiku’s price 90% below 100,000 tokens and lifted its computer-use score from 15.7% to 72.4%. Sierra and Meta published an OAuth-based way for an agent to prove who it is to a merchant. OpenAI replaced much of ChatGPT’s text output with clickable charts and calculators. Price, identity, interface — none of that is capability, and all of it is what stops an agent being deployed.

Two weeks ago the cuts were 40-50% at the frontier. This week’s is 90% at the cheap end, and it arrives with a small model that can actually drive a desktop. Last week agents got their own cloud machines and a free control plane. The stack is being filled in from the outside, one layer at a time.

The exception is the week’s one real capability claim. OpenAI posted 722 manuscripts covering 372 major maths problems and said nearly all of them came from a single prompt to a single agent — then withheld the prompts and the per-problem compute time. It is the most interesting thing announced this week and the least checkable.

What would confirm the plumbing read: the Personal Agent Protocol shipping a spec that Amazon, OpenAI or Anthropic actually sign. What would complicate it: Haiku 5.5 agents turning out to fail mostly in the long requests, where the price is five times higher.

---

Anthropic cuts Haiku to $0.10 per million input tokens

VentureBeat · 7 October 2026 — Haiku 5.5 costs $0.10 input and $0.50 output per million tokens on requests under 100,000 tokens, 90% below Haiku 4.5. Its OSWorld 2.1 computer-use score went from 15.7% to 72.4%, and Terminal-Bench 4.0 from zero to 39.2% at maximum effort. Above the threshold, rates are five times higher.

Why it matters: The cheapest tier is now good enough to run agent loops that needed a frontier model a month ago.

Sierra and Meta publish an identity standard for shopping agents

Forkast · 7 October 2026 — The Personal Agent Protocol, announced 6 October, extends OAuth so an agent can identify itself, declare intent and hold session-scoped permissions with a business. Stripe, Shopify, Walmart, Genesys, Instinct and Rocket are founding partners. Amazon, OpenAI and Anthropic are not, and no spec date was given.

Why it matters: Agentic commerce has had no identity layer at all; this is the first attempt with payments and retail behind it.

OpenAI dumps 722 maths manuscripts and keeps the prompts

Engadget · 7 October 2026 — OpenAI published 722 manuscripts on GitHub claiming progress on 372 major problems, including the four-dimensional Kakeya conjecture, from an unreleased “pioneer” model. The average result took about three hours of ChatGPT Pro. MIT’s Andrew Sutherland told Scientific American to “treat any claims about one-shotting problems with a single agent as unverified”.

Why it matters: A result nobody can reproduce is a marketing claim until the maths community says otherwise.

ChatGPT stops answering in prose

TechCrunch · 7 October 2026 — The GPT-6 rollout adds “Intelligent UI”: tappable buttons, task-specific calculators, editable charts and explanatory diagrams in place of much of the text. Pro, Plus, Business and Enterprise got it globally on Wednesday, Free and Go tiers on Thursday. Users can turn the visuals down.

Why it matters: If answers arrive as interfaces rather than paragraphs, every product sitting on top of a chat box needs rethinking.

---

The useful question going into next week is not which model is best. It is which one you can afford to run a thousand times before lunch.

Written by my AI assistant.