Daily 2026 · 06 · 29
EconomicTrend

Google rations Meta's Gemini access; Meta tells staff to economize on AI tokens

The Financial Times reports Google told Meta in March it could not deliver the Gemini capacity Meta was buying; Mark Zuckerberg's teams are now pivoting workloads to in-house Muse Spark and policing per-task token usage. Google Cloud's Q1 backlog has nearly doubled to $460 billion, several other clients have been capped to a lesser degree, and the move makes a peer-to-peer hyperscaler the first publicly disclosed customer to get put on a Gemini token diet.

CNBC · 2026-06-28 · Source ↗ ·

The Financial Times reported on Sunday that Google has been imposing limits on Meta's purchases of Gemini API capacity since around March, because Google Cloud cannot deliver the compute Meta is trying to buy. The disclosure landed in the FT's Sunday edition and was syndicated through CNBC, Bloomberg, Reuters and Dawn the same day. According to people the FT cites as familiar with the arrangement, the shortfall "disrupted and delayed some of Meta's internal AI projects," and Meta has told staff internally to be more efficient with the tokens they consume against Gemini. Other Google Cloud customers were affected to a lesser degree; Meta's exceptionally high demand made the squeeze especially acute.

The mechanics are mundane and that is exactly why the story matters. There is no contract dispute, no pricing fight, no platform-policy disagreement at the centre of it. Google looked at its forecast and told one of the world's biggest software companies that the GPUs needed to honour the order did not exist. The cap appears to have held for roughly three months by the time the FT got the story, which means Meta has been running at least one significant Gemini workload at a fraction of what its product teams asked for since the first quarter. Meta's response inside the building has been the closest thing the industry has yet seen to fuel rationing: a directive to engineers to economize per-task token usage, and a planned shift of critical workloads off Gemini and onto Meta's in-house Muse Spark model, the foundation system that Meta Superintelligence Labs has been standing up since the Alexandr Wang acquihire last summer.

Google has placed limits on Meta Platforms Inc.'s use of its Gemini artificial intelligence models because it could not provide as much computing capacity as the social media company wanted.— Bloomberg summary of the Financial Times report, June 28, 2026

Google Cloud's own Q1 2026 numbers are the right frame for reading the cap. Revenue grew 63 percent year over year to more than $20 billion in the quarter, and the cloud backlog — contracted future revenue not yet recognised — nearly doubled quarter over quarter to more than $460 billion. CEO Sundar Pichai told analysts on that call that compute constraints had prevented the cloud business from growing even faster. The Meta cap is the operational consequence of that constraint at its sharpest end. It is also the first time a hyperscaler has been publicly confirmed to be rationing access to a frontier model API to a peer hyperscaler, which is a different kind of fact than "we are running out of GPUs."

The disclosure sits inside a tight cluster of June 28 stories that all point at the same physical bottleneck. The Bank for International Settlements used its 2026 Annual Economic Report, released the same Sunday in Basel, to put an "AI bust" into the top tier of named risks to global financial stability, with hyperscaler debt and "circular financing" as exhibits. TechCrunch ran a Sunday feature calling Micron "the next Nvidia" after the memory maker's stock briefly cleared Meta and Tesla in market cap on a $41.45 billion fiscal Q3. Mark Gurman's Sunday Power On newsletter reported Apple and Microsoft both citing an "unprecedented" AI-driven memory shortage when justifying iPad, Mac and Xbox price hikes. Three different shapes of the same compute crunch landed on the same day; the FT's Google-Meta scoop is the operational one.

Meta's position in the picture is unusual. The company spent roughly $14 billion last year to bring in Alexandr Wang and a Scale AI cohort to anchor the Superintelligence Labs project, and it has committed somewhere between $115 billion and $135 billion in 2026 capex, with 7,000 workers reassigned to AI roles. The Muse Spark foundation model was built precisely so Meta would not depend on rivals' APIs for production workloads. The fact that Meta still wanted enough Gemini capacity to break Google Cloud's allocation means either Muse Spark is not yet at parity for the specific workloads in question, or Meta wanted Gemini for ablation and routing reasons even with parity. Either way, the public picture is that Meta is rationing tokens internally and migrating off Gemini at speed.

Reuters could not immediately verify the Financial Times report. Google and Meta did not respond to requests for comment outside business hours.— Reuters dispatch via Dawn, June 28, 2026

The downstream signal for everyone else is that the rationing logic is structural, not relational. If Meta — a hyperscaler customer in its own right, with a Cloud Backlog of its own and a direct corporate relationship with Google's top executives — cannot get the Gemini capacity it wants, second-tier customers should plan around the same constraint. That is part of why Sunday's Firmus / Nvidia announcement of a 360 MW DSX AI Factory in Batam, Indonesia, with 170,000 accelerators and $25 to $30 billion of projected offtake over six years, lands the way it does: it is incremental capacity aimed precisely at AI-native customers who cannot get a Google or Microsoft allocation. The competitive map for frontier inference is being redrawn around physical supply, not pricing or feature comparisons.

The cap also matters for how Google itself reports the next quarter. If Pichai again has to tell analysts that constraints capped growth, the Cloud number is going to be read through the FT story rather than the slide. Google paid SpaceX roughly $920 million a month earlier this year for "bridge" access to 110,000 Nvidia GPUs and stood up a wave of new sites; that bridge has not closed the gap fast enough to keep Meta at full draw. The fix is more buildout, and more buildout is the BIS's exact concern. The Gemini token diet at Meta is, in that frame, the first concrete piece of evidence that the demand the BIS is worried about is rationed today, not theoretical tomorrow.

Computing power constraints prevented even higher growth.— Sundar Pichai, Google CEO, on a recent Alphabet earnings call
Big Picture

Sunday was the day the compute crunch came out from behind the slides. The Financial Times reported Google has been rationing Meta's Gemini API access since March because Google Cloud cannot supply the capacity Meta is buying, and Meta is now policing per-task token use internally and migrating workloads onto its in-house Muse Spark. The Bank for International Settlements released its 2026 Annual Economic Report from Basel and put an "AI bust" into the top tier of named global risks, warning that the "circular financing" running between hyperscalers, chipmakers and neoclouds could be repriced into a 2008-scale credit event. Austrian Digital State Secretary Alexander Pröll wrote to EU Tech Commissioner Henna Virkkunen asking the bloc to pursue "the strategic establishment of Anthropic within the European Union," the first formal member-state intervention in the diplomatic fallout from the June 12 US export-control wall around Mythos 5 and Fable 5.

The infrastructure story carried the rest of the day. Sydney-based Firmus Technologies and Nvidia greenlit a 360-megawatt DSX AI Factory campus on Batam Island, Indonesia, under an eight-year deal that puts 170,000 accelerators through revenue-share and credit-support and projects $25 to $30 billion of offtake across the first six years, with the campus targeting a Q1 2027 launch and customers explicitly framed as AI-native firms that cannot get a hyperscaler allocation. TechCrunch's Kirsten Korosec ran a Sunday feature calling Micron "the next Nvidia" after the memory maker's stock briefly topped Meta and Tesla in market cap on a $41.45 billion fiscal Q3 and a $49 to $51 billion Q4 guide. Mark Gurman's Sunday Power On newsletter reported Apple and Microsoft both attributing iPad, Mac and Xbox price hikes to an "unprecedented" AI-driven memory shortage, and a Bloomberg cross-section found European investors hunting AI exposure in adjacencies after Sivers Semiconductors ran 2,246 percent year-to-date and STMicroelectronics ran 204 percent.

The dev-tooling layer carried light Sunday weight. ggml-org tagged six llama.cpp builds in a thirteen-hour window between b9829 and b9835, headlined by Kashif Rasul finally landing the long-pending DFlash v2 sliding-window-per-layer-types implementation as PR #22105 and ModelBest's 张涛 shipping a full MiniCPM5 chat and tool-call parser as PR #24889; Adrien Gallouët exposed the new `--offline` flag on the downloader CLI and Sigbjørn Skjæret added a `--dump-prog` Jinja template debugger. OpenAI Codex CLI cut an unannotated 0.143.0-alpha.29 pre-release at 00:30 UTC. Simon Willison surfaced Jon Udell's "Doctor, it hurts when agents create unreviewable PRs" and his reframing that "it's our loop, we recruit agents to join the team," and flagged DJ Patil's Hack Your Summer four-week sprint against the 2026 US student-internship freeze. Ford COO Kumar Galhotra and Vehicle Hardware VP Charles Poon told Bloomberg the company had rehired 350 veteran engineers after relying on AI quality systems alone had produced disappointing results, with Poon admitting Ford "mistakenly thought" AI ingesting design requirements would yield a high-quality product on its own.

More Today

11 stories · 2026-06-28
EconomicPolicy

BIS puts an "AI bust" into the top tier of global risks, warns "circular financing" could rival 2008

CNBC · 2026-06-28 · Source ↗

The Bank for International Settlements' 2026 Annual Economic Report lists an AI bust alongside inflation and sovereign debt as the gravest near-term risks to the global economy, with hyperscaler bonds and "poorly disclosed" equity-for-compute deals as exhibits.

The Bank for International Settlements released its 2026 Annual Economic Report at the institution's Annual General Meeting in Basel on Sunday, and used the report to elevate the durability of the AI investment cycle into the top tier of named global financial-stability risks. The report lists four pressure points the BIS now considers most alarming: re-accelerating inflation that risks un-anchoring expectations, lingering geopolitically driven supply shocks, financial fragilities tied to elevated asset valuations and investor complacency, and the durability of AI-related investment. It is the first BIS annual report to put AI capex into that tier of language.

The specific mechanic the BIS singles out is what it calls "circular financing." Chipmakers and hyperscalers take equity stakes in AI labs or neocloud providers; those labs and neoclouds then commit to multi-year purchases of chips and compute from the same companies that hold their equity. The terms of those agreements, the BIS writes, are "typically poorly disclosed," and the same physical assets can end up pledged multiple times across overlapping contracts. Hyperscaler debt issuance is the second exhibit: the report notes that Amazon, Alphabet, Microsoft, Meta and Oracle collectively issued more than $100 billion of corporate bonds in 2025, with a growing share routed through opaque off-balance-sheet structures, special-purpose vehicles, joint ventures and private credit funds.

Repricing of risk, whether triggered by higher interest rates or an AI bust, has the potential to be similarly disruptive to credit markets as the 2008 global financial crisis.— Bank for International Settlements, 2026 Annual Economic Report

The framing language goes further than past BIS reports. The institution writes that "supply bottlenecks and intense competition could lead to overinvestment seen in previous boom-and-bust cycles," and that disappointment in AI returns could trigger "a sudden pullback in financing and turn the capex boom into a protracted investment bust." The new risk the report names is a "sovereign-financial stability nexus," created by record-high public debt and a growing footprint of leveraged hedge funds in core bond markets, that would amplify a private-credit reversal into a sovereign one. BIS General Manager Pablo Hernandez de Cos used the AGM presentation to emphasise that "success depends on sound fiscal and financial foundations" and that "policy actions must reinforce each other to avoid a pull and push on the global economy."

The report's timing is sharp. It landed the same Sunday the Financial Times reported Google was capping Meta's Gemini access on physical capacity grounds and TechCrunch was calling Micron the next Nvidia after the memory maker's stock surged 236 percent in a month on a $41.45 billion fiscal Q3. The BIS frame treats those operational signals as exhibits in a structural argument: when the entity that places the equity check is also the entity that issues the forward order, the visible demand and the visible supply are the same balance sheet from two angles, and the audit trail that would catch a 2008-style cross-pledge problem is "poorly disclosed."

Guarantees embedded in some of these off-balance-sheet structures could activate unexpectedly.— BIS 2026 Annual Economic Report, on AI capex financing

The implications run in several directions at once. For central banks, the report is policy cover to push back on AI-driven leverage in the regulated banking system and to demand more granular disclosure from hyperscaler-affiliated lending vehicles. For Anthropic, which has filed for IPO, and OpenAI, which Bloomberg reported is eyeing a 2027 IPO, the report puts a yellow flag on the market backdrop those filings will land into. For neoclouds — CoreWeave, Lambda, Crusoe, Nebius, Together AI — the "circular financing" passage reads as a direct description of their capital stacks, and a warning that the next quarterly Bulletin from Basel will likely name names.

The BIS does not call for any specific action. The report is a risk-warning document and not a regulatory rulebook. But it does put on the official central-bank record, for the first time, that the durability of the AI investment cycle is a financial-stability issue rather than an investor-relations one. That distinction is what will get cited in the next round of Basel III review and the next set of country-level financial-stability reports. The audience inside the building heard de Cos describe the BIS's readiness to act if it observed an un-anchoring of inflation expectations; the audience outside the building heard the central bankers' central bank put the AI capex cycle on the same shelf as inflation and sovereign debt.

PolicyEconomic

Austria asks Brussels to actively recruit Anthropic onto European soil

U.S. News & World Report · 2026-06-28 · Source ↗

Austrian Digital State Secretary Alexander Pröll wrote to EU Tech Commissioner Henna Virkkunen asking the bloc to pursue "the strategic establishment of Anthropic within the European Union," the first formal member-state intervention prompted by the June 12 US export-control wall.

Austria has formally asked the European Union to consider hosting Anthropic inside the bloc as a strategic response to the June 12 US export-control directive that placed Anthropic's Claude Mythos 5 and Fable 5 behind US Commerce sign-off. The intervention came in a letter from Austrian State Secretary for Digitalization Alexander Pröll to European Commission Tech Vice-President Henna Virkkunen, reported Sunday by Bloomberg and syndicated through U.S. News, Business Recorder and MarketScreener. It is the first publicly disclosed formal member-state request to relocate a US frontier lab in response to the Trump administration's trusted-partner regime.

Pröll's framing in the letter is explicit. He asked Virkkunen to explore "the strategic establishment and participation of Anthropic within the European Union. With legal certainty, market access, capital and a set of values that suits this company." He added that it was important Europe was not cut off from major innovations and acknowledged scepticism about how the plan would work in practice. The letter does not propose specific mechanics — equity investment, tax structure, regulatory carve-out, or sovereign compute commitment — and Bloomberg notes Pröll himself flagged that feasibility is uncertain. The diplomatic act is the request itself.

The strategic establishment and participation of Anthropic within the European Union. With legal certainty, market access, capital and a set of values that suits this company.— Alexander Pröll, Austrian State Secretary for Digitalization, in his letter to the European Commission

The context is the rolling consequence of the June 12 US Commerce directive that pulled Anthropic's Mythos 5 and Fable 5 from all global customers, and the June 26 Lutnick letter that partially restored Mythos 5 access for roughly 100 institutions named in Annex A of the order. Fable 5 remains globally offline. Foreign nationals can access Mythos 5 only when they work for a listed institution. Allied jurisdictions — the UK, Japan, the rest of Five Eyes, and the EU — are not on Annex A. Pröll's letter is the first formal articulation that some EU governments are now willing to think about pulling the lab itself onto European soil rather than waiting indefinitely for the export-control regime to be widened.

The implications stretch beyond Austria. If even one EU member state is asking Brussels to pursue Anthropic relocation, that puts the question on the Commission's table in a way that the Sakana Fugu launch and the 360 Tulongfeng release earlier in the week — both regional alternatives positioned for customers locked out of Mythos 5 — did not. It also opens a counter-conversation in Washington about whether export-control authority used as an industrial-policy instrument has the effect of pushing allied governments to negotiate around it rather than through it. The Lutnick letter explicitly named "trusted partners" without listing any non-US entities; Pröll's letter is the first formal response to that omission from a friendly capital.

Anthropic has not commented on the letter publicly. The company's Series H at a $965 billion post-money valuation was underwritten in part by Mythos 5 demand from US Fortune 500 cyberdefence teams, and the partial restoration restored that revenue line. A formal EU pursuit of relocation would create awkward questions about corporate structure, sovereign data residency, and the relationship between a US-headquartered lab and the European executive branch; in the worst case for Washington, it would create the template by which the EU offers to host any US lab the next administration restricts. For Brussels, Pröll's letter is the first concrete signal that the "digital sovereignty" rhetoric of the past several Commissions has an interlocutor on the corporate side.

It was important Europe was not cut off from major innovations.— Alexander Pröll, paraphrased by Bloomberg, June 28, 2026

The Pröll request also reframes the diplomatic register of the June 12 directive. Before Sunday, the export-control story was about US lab access and US trusted-partner gating. The Austrian letter inserts a third axis: the corporate base. The question on the European table is no longer whether the Commission can secure EU access to Mythos 5; it is whether the Commission should be in the business of recruiting Anthropic itself. The answer Virkkunen gives back will set the tone for how every subsequent restriction lands inside Brussels, and how every other EU member state writes its own letter.

EconomicApplication

Firmus and Nvidia greenlight 360 MW Batam AI factory, 170,000 accelerators, $30 billion expected offtake

The Next Web · 2026-06-28 · Source ↗

Sydney-based Firmus Technologies and Singapore's DayOne signed an eight-year revenue-share with Nvidia for a 360-megawatt DSX AI Factory on Batam Island, targeting Q1 2027 launch and $25 to $30 billion of committed offtake in the first six years.

Firmus Technologies, the Sydney-based AI infrastructure company valued at $5.5 billion post-money, signed an eight-year partnership with Nvidia and Singapore-based DayOne on Sunday to build a 360-megawatt Nvidia DSX AI Factory campus on Batam Island, Indonesia. Bloomberg broke the story Sunday morning; The Next Web confirmed terms with a 17:16 UTC dispatch and Reuters distributed a Scott Murdoch dispatch through the KELO mirror with the same details. The deal gives Firmus access to up to 170,000 Nvidia AI accelerators delivered in batches from Q1 2027 through early 2028, structured as revenue-sharing plus credit support rather than a flat capacity contract.

The commercial pitch is unusually concrete for a buildout announcement. Firmus and DayOne expect $25 to $30 billion in committed offtake agreements during the first six years of the campus's operation. The customer profile is AI-native companies — frontier labs, model providers, and second-tier AI firms — that have been priced out of, or rationed by, US hyperscaler capacity. That framing is consistent with the Financial Times Google-Meta story that landed the same day: when even Meta cannot get the Gemini capacity it wants from Google Cloud, the relevant audience for an AI-native infrastructure provider is everyone whose order is one tier smaller than Meta's.

We have worked to figure out how to close the gap between the cost benefits that the large guys have access to, and the guys that are up and comers.— Tim Rosenfield, Firmus Co-CEO

The Firmus capital stack is worth reading next to the order. The company raised $1.35 billion over the past six months, including a $505 million April 2026 round led by Coatue Management with Nvidia among the backers. Bloomberg reports Firmus is expected to pursue an IPO later this year, with investment banks already appointed for the preparation. Founded in 2019 as a Tasmania-based Bitcoin mining operation, the company pivoted to AI infrastructure in the past eighteen months. The Batam campus is its first Indonesian project; the existing Australian pipeline is structured around hyperscaler customers and a separate 1.6-gigawatt commitment with CDC Data Centers across Australia targeting 2028.

The geography is part of the pitch. Batam is the Indonesian island directly across the Singapore Strait from Singapore, which puts the campus inside the Southeast Asian financial centre's connectivity envelope while keeping land, power and labour costs Indonesian rather than Singaporean. That is the same play DayOne's parent company has been running for the past several years; Firmus has now anchored a multi-tenant AI-native build on top of it. The benchmark in the Asia-Pacific AI build cycle is Blackstone-backed AirTrunk's $30 billion India commitment, announced earlier this year; Firmus and DayOne's Batam announcement is the same dollar magnitude landed in a different jurisdiction and aimed at a different customer set.

We're building our business based on demand that we're seeing from customers and contracts that we're closing.— Tim Rosenfield, Firmus Co-CEO, responding to AI-bubble concerns

The structure also matters. The revenue-sharing component aligns Nvidia's incentives with utilisation rather than just sell-through — Nvidia gets product revenue plus a share of cloud revenue — and the credit-support component means Firmus does not have to fund the full 170,000-GPU outlay from its own balance sheet ahead of Q1 2027 deliveries. That is the same structural pattern the BIS report released the same Sunday flagged as "circular financing": equity holder is also forward customer, terms not fully disclosed, asset pledged into multiple downstream contracts. Firmus has not disclosed the full revenue split, which is consistent with the BIS's broader observation that these agreements are "typically poorly disclosed."

For the regional frontier-model picture, the announcement is the second piece of Southeast Asian capacity for AI-native customers to land in two weeks, alongside Sakana's Fugu launch and a wave of Chinese frontier-model releases. A campus that opens in Q1 2027 with 170,000 accelerators changes the negotiating position of every second-tier model lab that wants to train or serve a frontier-scale system in the region without going through US export controls. The Firmus pitch is that those labs no longer have to choose between waiting on a hyperscaler allocation and downscaling their ambitions; Batam is now an option on the menu.

EconomicTech

Wall Street calls Micron "the next Nvidia" after $41.45B fiscal Q3 and a 236-percent monthly run

TechCrunch · 2026-06-28 · Source ↗

Micron's memory boom briefly cleared the stock past Meta and Tesla in market cap on Thursday; TechCrunch's Sunday feature collects 16 signed Strategic Customer Agreements including Anthropic and Nvidia and a William Blair Outperform call.

TechCrunch's Kirsten Korosec ran a Sunday feature on the Boise-based memory maker Micron under the headline "Why Wall Street thinks US memory maker Micron is the next Nvidia," collecting the past six weeks of share-price action, the fiscal Q3 print, and the long-term supply agreements that the company has signed against the AI compute cycle. The piece was filed at 8:00 a.m. Pacific on June 28 and is the cleanest summary to date of why Micron has come to function as the memory leg of the same trade that runs through Nvidia.

The numbers carry the story. Q3 revenue was $41.45 billion, roughly quadrupled year over year. Profit was $28.2 billion, up from $1.88 billion in the prior-year quarter. Q4 guidance is $49 to $51 billion, with the upper bound implying another sequential leg up. The stock closed Friday at $1,132 a share, having run 236 percent in a single month. Thursday's intraday print briefly took Micron's market cap past Meta and Tesla — the first time the memory maker has cleared either name on a market-cap basis — and put the company on a roughly $1.27 trillion line going into the weekend.

Given the strong likelihood of continued ASP growth in the coming quarters and improving revenue visibility thanks to a rapidly expanding set of long-term agreements with key customers, we see potential for more durable earnings growth and reiterate our Outperform rating.— Sebastien Naji, William Blair memory analyst

The structural change behind the numbers is the long-term supply agreement layer. Micron has now signed 16 Strategic Customer Agreements across data-centre, consumer and automotive segments, designed to insulate the company from the historic memory boom-bust cycle that always cratered earnings during DRAM gluts. The data-centre roster cited in the piece runs across Microsoft, Amazon Web Services, Google, Meta, Oracle and Nvidia at the system level, and the Anthropic SCA — formalised on June 26 in the Series H pact reported in the previous Apollo edition — is the marquee frontier-lab name. The agreements are multi-year, with embedded HBM, DRAM and SSD volume commitments that map directly to the next two generations of frontier-model training and inference clusters.

Korosec uses the term "RAMageddon" for the memory shortage driving the move, and notes that the analyst consensus now expects the tightness to persist into 2027. The Sunday feature pairs with two other June 28 datapoints in the same physical-supply cluster: the Financial Times Google-Meta scoop on Gemini rationing, and Mark Gurman's Sunday Power On newsletter, which reports Apple and Microsoft hiking iPad, Mac and Xbox prices with explicit reference to the AI-driven memory shortage. The shortage that capped a Meta Gemini order is the same shortage that just made Micron a trillion-dollar company.

The strategic agreement with Anthropic to scale HBM supply was a milestone in shifting Micron from a cyclical memory commodity name to a frontier-AI infrastructure supplier.— TechCrunch summary of the Anthropic SCA, June 28, 2026

The uncertainty Korosec flags is the same one Wall Street has been chewing on for the past month: whether Micron can hold the profitability without the cyclical reset. The SCAs are the structural answer — locked-in customers do not flush inventories at the bottom of a downturn the way merchant-market customers do — but the SCAs only cover the data-centre and HBM volumes, not the consumer DRAM that historically dominated the cycle. The TechCrunch piece does not resolve the question; it lays out the bull case in the analyst voice and notes the bear case in the structural one.

For the AI-compute story, the Micron print is the supply-side mirror of the Google-Meta cap and the Firmus-Nvidia Batam build. Frontier inference is rationed today because the HBM, DRAM and accelerator stack at the back of the cluster cannot scale at the pace the demand is moving. Micron is the company building the back of the cluster, and its order book is now what the equity market is reading as evidence that the supply curve is going to take years to catch up. The Anthropic and Nvidia SCAs are the largest single contracts under that bet.

EconomicTrend

Apple and Microsoft hike iPad, Mac and Xbox prices, both blaming an "unprecedented" AI memory shortage

Bloomberg Power On · 2026-06-28 · Source ↗

Mark Gurman's Sunday Power On dispatch frames the simultaneous Apple and Microsoft price hikes as "the AI era coming home," with both companies citing memory-chip scarcity as the explicit cause and previewing the M6, M7 and Touch MacBook Pro roadmap.

Bloomberg's Mark Gurman used his Sunday Power On newsletter to assemble the past week's consumer-electronics price moves into a single narrative. Apple raised prices across its iPad and Mac lines; Microsoft raised prices on the Xbox console line. Both companies attributed the moves to a memory-chip shortage that both characterised as unprecedented, and both pointed at the AI demand cycle as the upstream cause. The Sunday dispatch frames the simultaneous moves as the moment the AI compute crunch landed inside the household budget.

The mechanic is the same one Micron just printed against. AI-driven HBM and DRAM demand has pulled supply out of the consumer memory market and pushed average selling prices up across the board. Apple and Microsoft both have multi-year supply contracts that absorb some of the move, but the magnitude this cycle has cleared what the contracts can dampen, and both companies have chosen to pass the cost through rather than compress margin. Gurman's previous reporting put Micron DRAM average selling price up more than 260 percent and the stock up 800 percent year-on-year; the Sunday piece pulls those numbers into the consumer frame for the first time.

Apple's sweeping price hikes bring the AI era home.— Mark Gurman, Power On, June 28, 2026

The Apple roadmap that lives behind the price moves is the second piece of Gurman's Sunday dispatch. The newsletter previews the M6 and M7 silicon generations and the Touch MacBook Pro that Apple is positioning as the answer to the AI-PC category that Qualcomm, Intel and AMD have all been pushing for the past year. The price hikes set the expected entry price for the M6 and M7 generations higher than the M5 line, which is the first time in five years that an Apple silicon transition has cleared the prior generation's pricing instead of holding it or undercutting it.

Microsoft's Xbox move is the more revealing of the two for the industry frame, because the console business has historically been the part of consumer electronics where memory cost was the most managed input. If Microsoft has to raise Xbox prices on memory grounds, that is the strongest possible signal that the AI demand cycle has cleared every consumer-electronics input it was going to clear. The Apple price moves can be read as Apple optimising margin against scarcity; the Microsoft Xbox move can only be read as the cycle reaching a category where price moves are last resort.

The memory-chip shortage that has driven up the cost of memory chips by an unprecedented amount has gotten so dire that both Apple and Microsoft are pointing to it as the reason for their respective price hikes.— Bloomberg framing of the Apple and Microsoft statements, June 28, 2026

The downstream signal is straightforward. Every consumer device with a meaningful memory footprint is now on notice that the next generation is going to clear at a higher price than the current one until the HBM and DRAM supply curve catches up to data-centre demand. The TechCrunch Micron piece, the Power On Apple piece, the Bloomberg Microsoft piece and the FT Google-Meta piece all land on the same Sunday because they are all the same trade: AI capex demand is pulling the global memory supply curve and the consumer electronics market is the second-order effect.

EconomicTrend

European AI money rotates into power utilities and optical specialists; Sivers up 2,246 percent year-to-date

Bloomberg · 2026-06-28 · Source ↗

A Bloomberg cross-section finds European investors bidding up Sivers Semiconductors, STMicroelectronics and Nokia as proxy AI trades, on the theory that the Continent has too few direct names and the capex cycle is fundamentally a power and grid story.

Bloomberg ran a Sunday survey of where European investors are putting their AI exposure now that the most-direct US and Asian names are crowded and expensive. The answer the piece sketches is that the Continental money has rotated into adjacencies — power utilities, banks with AI-lending exposure, and specialist semiconductor and networking names where European players still have a defensible position — rather than into a thin handful of European AI pure-plays that do not really exist at scale.

The specific names the piece highlights tell the story. Sivers Semiconductors AB, the Swedish maker of laser arrays and optical engines that move data as light inside AI data-centre interconnects, is up 2,245.93 percent year to date in 2026 and is now the best-performing stock on any major European exchange this year. STMicroelectronics, which makes the power chips, microcontrollers, sensors and increasingly the optical products that link AI servers, is up 204.28 percent and has raised its 2026 data-centre revenue ambition to roughly $1 billion, up from a prior guide of "nicely above $500 million." Nokia is in the middle of reinventing itself as a supplier of networking and optical kit for AI data centres after years as a primarily telecoms-equipment vendor; the share has tracked the same data-centre-build cycle Bloomberg is pricing.

Europe's hunt for AI stocks leads to power suppliers and banks.— Bloomberg headline, June 28, 2026

The structural read is the second half of the piece. The reason European investors are buying utilities and banks rather than model labs is that the AI capex cycle, on the Continent at least, is being priced as an electricity and grid story rather than a software one. The data-centre builds that anchor frontier inference need power-purchase agreements, transmission upgrades, and project finance, and European utilities and banks are the participants on the supply side of all three. The model layer at the top of the stack is overwhelmingly US-headquartered; the European exposure is one layer down, in the physical infrastructure and the financing that powers and pays for it.

Sivers' year-to-date number is the cleanest illustration of how concentrated the proxy trade has become. A 22-bagger in six months reflects a thin float meeting a wall of allocation flow from investors who want some piece of the data-centre interconnect cycle. STMicroelectronics' smaller but still triple-digit move reflects the same flow at a more liquid name. Both moves carry the same risk as the broader BIS warning that landed the same Sunday: if the AI capex cycle reprices, the proxies reprice with it, and the proxies have less volume to absorb a sell-side rotation than the direct US names do.

Sivers' lasers and optical engines move data as light inside AI data centres, and the stock has become the cleanest European proxy for the data-centre interconnect cycle.— Bloomberg paraphrase, June 28, 2026

The piece also flags a second-order question for Brussels and London. If the easiest European exposure to the AI cycle is the Continental power and grid layer, then the regulators governing electricity transmission and project finance have become inadvertent AI regulators. The Austrian letter asking Brussels to recruit Anthropic, also reported Sunday, is the first sign that European member states are beginning to think of AI policy as industrial policy in those terms; the Bloomberg piece supplies the equity-market reading of the same shift.

ApplicationTrend

Ford rehires 350 "gray-beard" engineers after AI quality systems fall short

TechCrunch / Bloomberg · 2026-06-28 · Source ↗

COO Kumar Galhotra and Vehicle Hardware VP Charles Poon told Bloomberg the company "mistakenly thought" AI ingesting design requirements would deliver a high-quality product on its own; 350 veterans have been rehired to retrain the models and the staff using them.

Ford rehired roughly 350 veteran engineers — some former Ford employees and others pulled from suppliers — to retrain younger staff and to reprogram the company's automated quality systems after relying on AI-driven quality processes alone produced what executives described as disappointing results. Bloomberg broke the story Friday and TechCrunch's Anthony Ha distributed an aggregated version Sunday at 12:05 p.m. Pacific. The substantive admission inside the piece is the closest a major US manufacturer has come to publicly walking back an AI-replacement strategy in favour of AI-augmentation.

The mechanics matter. Ford had pushed automated quality systems into more of the production process over the past several years on the theory that the systems could ingest design requirements directly and catch defects that human engineers had been catching manually. Vehicle Hardware Engineering Vice-President Charles Poon told Bloomberg the bet did not work. Chief Operating Officer Kumar Galhotra said the company had been "relying more and more on automated quality systems" and that the results had been disappointing enough to justify bringing the veterans back. The rehires are positioned as a retraining cohort — they teach the younger engineers what to look for, and they help reprogram the AI systems with the patterns the AI was missing.

Mistakenly we thought that by just introducing artificial intelligence and ingesting the design requirements that we had, that that would produce a high-quality product.— Charles Poon, Ford Vice-President of Vehicle Hardware Engineering

Ford is not abandoning AI. The framing the executives use is exactly the augmentation frame the Jon Udell post Simon Willison surfaced the same day argued for in software engineering: it is our loop, the agents join the team. The rehires are there to recruit the AI back into a workflow it had been put in charge of. The company anticipates roughly $1 billion in cost reductions this year on the broader programme and has just claimed the top spot among mainstream brands in the JD Power Initial Quality Survey, which gives Galhotra cover to talk about the walkback without it reading as a strategic retreat from automation.

The story is the first concrete data point in the manufacturing wing of the AI-replaces-people vs AI-augments-people debate that has dominated the productivity discourse all year. The numbers are large enough to matter — 350 rehires is not a pilot — and the executives are senior enough that the admission cannot be dismissed as middle-management excuse-making. The wording Poon chose — "mistakenly we thought" — is a deliberate, on-the-record reversal of the previous strategy.

Relying more and more on automated quality systems… brought back technical specialists to identify the failure points.— Kumar Galhotra, Ford Chief Operating Officer, paraphrased by Bloomberg

The implications for the broader AI-in-industry story run two ways. The first is that other manufacturers — General Motors, Stellantis, Volkswagen, Toyota — will now read Ford's walkback as cover to make similar walkbacks in their own production lines, which means the next several quarters of capital-allocation reporting from auto OEMs will likely contain a higher proportion of "human-in-the-loop" framing than the prior several did. The second is that the Goldman Sachs estimate of 11,000 US jobs erased per month by AI, which circulated heavily earlier this month, has to be read against the possibility that some of those erasures will need to be reversed at scale. Ford is the first major employer to do the reversal publicly.

Tech

llama.cpp tags six Sunday builds; DFlash v2 sliding-window-per-layer-types lands, MiniCPM5 tool-call parser ships

ggml-org/llama.cpp · 2026-06-28 · Source ↗

Six tags between 06:46 and 19:37 UTC, headlined by Kashif Rasul landing the long-pending DFlash v2 architecture, ModelBest's 张涛 shipping a full MiniCPM5 chat and tool-call parser, and Adrien Gallouët exposing `--offline` on the downloader CLI.

ggml-org tagged six llama.cpp builds in a thirteen-hour Sunday window, b9829 through b9835 with b9834 skipped. The headline pieces are an architectural addition and a chat-template parser; the rest is housekeeping. The thirteen-hour drumbeat is unusual for a weekend day and reads as the project clearing a backlog of merged-but-untagged work ahead of the start of the work week.

DFlash v2: sliding-window-per-layer-types finally upstream

The 13:38 UTC b9831 tag landed Kashif Rasul's PR #22105, which has been open and reviewed for months. The implementation adds DFlash v2 support and implements sliding-window attention on a per-layer-types basis rather than a single global window, matching what the v2 architecture specifies. The practical effect is that downstream open-weights releases that have been built against the v2 sliding-window pattern — Mistral Small 4, Qwen3, Gemma 4 in their long-context variants — can now light up in the upstream llama.cpp graph without bespoke kernels. The PR also adds the architecture documentation section, which is how downstream quantisers know which kernel paths to target.

MiniCPM5: ModelBest gets first-class tool calls

The 15:32 UTC b9833 tag landed 张涛's PR #24889, which implements the MiniCPM5 chat parser and a full tool-call parser. The PR went through several review rounds — a PEG parser refactor on reviewer feedback, a Jinja min/max API compatibility fix against upstream Jinja2, an autoparser for XML-format tool calls, a streaming tool-arg placeholder handling fix, a reversion of shared mapper fallbacks per review, and added message delimiters and grammar simplifications. The result is that MiniCPM5 — ModelBest's small-model series that is one of the most-downloaded SLM lines for on-device agents — now has first-class function-calling support in llama.cpp, narrowing the gap with vLLM and SGLang for on-device agent stacks.

chat: implement minicpm5 parser.— PR #24889 commit message, ModelBest

The housekeeping layer

The 06:46 UTC b9829 tag is logging verbosity reduction across server and common modules under PR #25078, plus an internal rename of the CMN_ prefix to COM_. The 11:03 UTC b9830 tag is Adrien Gallouët's PR #25091, which exposes the `--offline` flag to the new `llama download` CLI so users can force offline behaviour on downloads as well as inference startup, and fixes a use-after-free in the URL-task callback path that powers Hugging Face and GGUF downloads. The 14:18 UTC b9832 tag is Sigbjørn Skjæret's PR #25086, which adds a `--dump-prog` flag for debugging Jinja chat templates by dumping the compiled template program for inspection. The 19:37 UTC b9835 tag is PR #25084, a UI fix for the stop button and reasoning-token skip controls in the new single-model mode of the built-in WebUI.

The Codex CLI also moved Sunday but in a smaller way. OpenAI's 0.143.0-alpha.29 pre-release was cut at 00:30:41 UTC by the github-actions[bot] account, ships the full 146-asset npm and symbol bundle across darwin, linux and windows on aarch64 and x86_64, and carries the literal body text "Release 0.143.0-alpha.29" with no published changelog. The prior week's 0.142.3 remains the stable, and alpha.29 is a routine bump in the 0.143 alpha track rather than a feature drop.

For the broader inference and agent stack, the consequential piece is DFlash v2. Sliding-window-per-layer-types is exactly the architectural primitive the next wave of long-context open-weights releases needs, and having it in the upstream graph means the quantised variants on Hugging Face will be usable on commodity inference stacks within days of release. The MiniCPM5 parser is the bigger immediate-impact change for the on-device agent community; the two together make this Sunday's drumbeat one of the more substantive single-day llama.cpp tag clusters in recent weeks.

Tech

Simon Willison amplifies Jon Udell's "it's our loop, we recruit agents to join the team"

Simon Willison's Weblog · 2026-06-28 · Source ↗

Willison signal-boosts Jon Udell's "Doctor, it hurts when agents create unreviewable PRs" and the reframing it argues for: the human is not in the agent's loop, the agent has joined the human's team.

Simon Willison posted a short curation entry on Sunday at 21:57 UTC pointing at Jon Udell's blog post titled "Doctor, it hurts when agents create unreviewable PRs. Don't do that." The Willison post is brief; the value is the signal-boost and the framing he adds to it. Udell's argument is that the phrase "human in the loop," which has become the standard liability-shielding framing for any deployment of generative AI in a workflow, implicitly hands authority to the machine and casts the human as a passive validator inside a process the machine owns. Udell flips it: the workflow is the human's, the agents are recruited into it as collaborators, the loop is ours and not theirs.

It's our loop, we work the same way we always have, now we recruit agents to join the team.— Jon Udell, paraphrased on Simon Willison's Weblog

The specific pain Udell names is the agent that produces a pull request too large or too entangled for a human reviewer to actually read. The unreviewable PR is the failure mode where the agent has, in practice, taken authority away from the team it was nominally serving, because the team cannot meaningfully veto what the agent wrote. The Udell rule that follows is "don't do that" — instruct the agent to ship changes at the size and scope a human can read and reason about, and treat the reviewer's bandwidth as the constraint the agent has to respect, not the other way around.

The Willison amplification is itself the news. Willison is one of the most-read voices on coding-agent practice, and his curation choices function as a sorting layer for what working engineers should be reading. By putting Udell's reframing in front of his audience the same day Ford's Charles Poon admitted that thinking AI alone could deliver quality was a mistake, Willison joined the two pieces into a single thesis: AI augments teams, it does not run them. The Bloomberg and TechCrunch pieces on Ford are the manufacturing-floor version of the same argument the Udell post is making for software.

The post is tagged under AI, generative AI, LLMs, coding agents, and agentic engineering. Willison's catalogue of these tags is now itself a curated reading list, and the inclusion of the Udell post under all five tags signals that he expects this reframing to be load-bearing for how working teams talk about agent integration over the next several quarters.

TechApplication

DJ Patil's "Hack Your Summer" opens July 13 cohort against the 2026 US student internship freeze

Simon Willison's Weblog · 2026-06-28 · Source ↗

A four-week production sprint for undergrads, grad students and recent grads, framed as a response to the US company-hiring freezes that hollowed out the 2026 summer internship calendar; application deadline July 8, cohort begins July 13.

Simon Willison flagged DJ Patil's Hack Your Summer programme on his blog Sunday at 19:26 UTC, sourcing the item from Patil himself. The programme is a four-week intensive sprint for undergraduates, graduate students and recent graduates, framed explicitly as a response to the 2026 US student-internship shortage that has resulted from the past year of company hiring freezes — most of which have been justified in earnings-call language as freezes triggered by AI-driven productivity gains.

The structure is a high-velocity production sprint. Participants build tangible, public-facing work over four weeks, with mentor and peer guidance built into the cohort. The second cohort, which is free, begins on July 13. The application deadline is July 8. The programme is also recruiting mentor volunteers, and Patil's pitch positions the programme as a way for established practitioners to take a direct hand in keeping the next cohort of engineers in the talent pipeline through a season when the institutional hiring engines have stalled.

Hack Your Summer is a four-week production sprint to give students who can't find a 2026 internship something real to build, ship and put on a resume.— Programme framing, surfaced via Simon Willison's Weblog

The Willison post is short and operational rather than analytical. The relevance for the AI-news beat is the inverse-shape relationship the programme has with the broader headline. The same year the Goldman Sachs estimate of 11,000 US jobs erased per month by AI dominated the productivity discourse and Ford was rehiring 350 veterans to fix what AI quality systems alone could not, a privately-organised four-week sprint is being stood up to keep students moving through the early-career pipeline that company internships used to feed. Patil is one of the original architects of US federal data-science capacity; the choice of Hack Your Summer as the response, rather than a policy intervention, is itself a comment on where the practitioner community thinks the levers actually are right now.

Tech

OpenAI Codex CLI cuts unannotated 0.143.0-alpha.29 at 00:30 UTC

openai/codex · 2026-06-28 · Source ↗

A routine CI bump in the active 0.143 alpha track: full 146-asset cross-platform bundle, body text literally "Release 0.143.0-alpha.29," no changelog. The prior week's 0.142.3 remains the stable.

OpenAI's Codex CLI cut pre-release 0.143.0-alpha.29 at 00:30:41 UTC on Sunday, tagged by the github-actions[bot] account against commit e4198a095c36a9f8f703b61be900f66df85e0984. The release ships the full 146-asset cross-platform bundle — npm packages, symbol bundles, and binaries for darwin, linux and windows on aarch64 and x86_64 — and carries the literal body text "Release 0.143.0-alpha.29" with no published changelog. The page has picked up five emoji reactions, one laugh, three rockets and one eyes, which is the working community's shorthand for "noted, awaiting notes."

The pattern reads as a routine CI bump rather than a feature drop. The prior week's 0.142.3 remains the stable release, and the 0.143 alpha track is the active pre-release stream against which the next stable will eventually land. There is no public dependency change, no advertised behavioural shift, and no security-related note. The release exists because the alpha train is set to cut on commit cadence rather than narrative cadence; the substantive 0.143 changes will appear in a later alpha or in the eventual 0.143 stable.

The reason a no-notes alpha is worth a line is that it confirms OpenAI is continuing to push the Codex CLI alpha train through weekend windows without gating on changelog publication. For a working agent-CLI user base, the practical implication is that anyone tracking the alpha channel has to pull and test rather than read; for a tooling-stability audience, the implication is that the gap between 0.142.3 and the next stable will likely be wider than the alpha cadence suggests.