Plutonic Rainbows

Meta is cutting 8,000 jobs starting May 20, roughly ten percent of its workforce, and not filling another 6,000 open roles. Microsoft is offering early retirement and voluntary buyouts to about 8,750 US employees, near seven percent of its US headcount. Both announcements landed inside the same 48-hour window. Both companies report earnings on Wednesday. And both are on track to spend a combined sum on AI infrastructure this year that makes those payroll numbers look like rounding.

The framing the companies are using is interesting because it has finally stopped pretending. Janelle Gale, Meta's chief people officer, told staff the cuts are needed to "offset the other investments we're making." She did not say which investments. She did not have to. Zuckerberg has said the quiet part out loud several times this year already: 2026 is the year of "major AI acceleration," with planned spend north of $115 billion on data centres, custom silicon, and the people who can build them. Microsoft's number is comparable. Across Meta, Microsoft, Alphabet, and Amazon the collective AI infrastructure outlay this year clears $700 billion.

The arithmetic, then, is plain. You move money from one column to another. You stop paying a recruiter and start paying for a GB200 rack. You stop paying a layer of middle management at Reality Labs and start paying TSMC for a wafer allocation. The headcount line shrinks because the capex line is eating it. There is no hidden mystery about where the money is going.

What's new is that the labour story is no longer being told through the language of a downturn. There is no recession to blame. Hyperscaler revenue is up. The cuts are not because business is bad. They are because the business has decided that a particular shape of human labour is now optional. Coding, recruiting, ops, mid-tier programme management. The kinds of work where an agent does eighty percent of the task and a smaller team does the cleanup.

I am not sure the agents are quite there yet. The shootouts between GPT-5.5 and Opus 4.7 are still close enough that nobody outside the labs can confidently call a winner on a given task, and the public benchmarks have a known habit of flattering the model that wrote them. But the executives are not waiting for proof. They are pricing the bet now, against this year's salary budget, on the assumption that the gap closes before the next fiscal year begins. If it does, the cuts look prescient. If it does not, the cuts still happened, and the people are still gone, and the inference bill arrives anyway.

There is an honesty to it that I almost respect. For most of the last decade, "efficiency" was the euphemism that companies reached for when they wanted to fire people without saying why. The word still gets used. But the underlying bookkeeping has shifted. Efficiency now means a specific trade: a payroll line exchanged for a compute line, a headcount slot exchanged for a token bill that the CFO can model with a straight face.

Wednesday's earnings will be the test. If the analysts ask about the labour impact in the same breath as they ask about Azure growth, the equation has been accepted. If they ask only about capex, it already has been.

OpenAI shipped GPT-5.5 on Thursday. The model is available now to paying ChatGPT and Codex users in standard, thinking, and pro flavours, with API access to follow. The pitch from Greg Brockman in the press briefing was that this one is built for work, coding, computer use, research, and that it can take an unclear problem and decide what needs to happen next without much hand-holding.

Stripped of the briefing-room varnish, the message is that ChatGPT, Codex, and the browser tooling are converging into a single product, and 5.5 is the engine that makes the convergence plausible. Brockman called it the foundation for "how we're going to do computer work going forward." That is super-app language, and it has been the open secret of OpenAI's product strategy for about a year. The model release is the part that gets the headlines; the strategy is the part that decides whether the next twelve months go well.

I am sympathetic to the ambition. The frustration of using six different AI surfaces to get one task done is real, and the seam-stitching tax adds up. A single thing that opens your browser, edits your repo, runs your tests, and writes the PR description is genuinely useful, more useful than another point on a benchmark. The hard part has never been the demo. The hard part has always been getting the model to know when to stop, when to ask, when to fail loudly rather than quietly produce something broken.

5.5 is priced at $5 per million input tokens and $30 per million output, which is GPT-5 territory and roughly an order of magnitude above V4 Pro. That is fine if the agentic capability is genuinely a step up, and a problem if it is not. Computer-use agents burn output tokens prodigiously. A single half-decent coding session can produce tens of thousands of tokens of tool calls, reasoning traces, and revisions. Multiply that by an enterprise rollout and the unit economics get scary fast, particularly when a Chinese open-weight model can run the same loop, less well, for pennies.

The other thing worth noticing is the cadence. Anthropic shipped Opus 4.7 the week before. DeepSeek previewed V4 the day after. CNET wrote it up as an "arms race," which is the laziest possible framing but, this week at least, accurate. Three frontier releases inside eight days, all pitching some flavour of agentic coding as the headline capability, all aiming at the same enterprise budget. The dispersion of "best at coding" across labs keeps narrowing. So does the differentiation.

Tom's Guide ran a 5.5 versus Opus 4.7 head-to-head and reported seven wins for Claude on seven impossible tasks. Single-evaluator shootouts are noise more than signal, but the noise is itself informative: nobody outside the labs is sure which model wins on which kind of work right now, and the customer-side answer is increasingly "whichever one we already have a contract with." That is not where OpenAI wants to be when it is asking for super-app trust.

The model probably is good. The strategy probably needs more than a model.

Cohere announced on Friday that it would merge with Aleph Alpha, the German enterprise AI company that pivoted away from frontier model development in 2024. The combined entity will be valued at roughly $20 billion once Cohere's pending Series E closes, with Schwarz Group, the German retail conglomerate that already co-led Aleph Alpha's 2023 Series B, putting in another $600 million. Dual headquarters in Toronto and Heidelberg. Aidan Gomez stays at the helm. The press release uses the phrase "transatlantic AI powerhouse" without flinching.

The strategic logic is more interesting than the dollar figure. Cohere was last valued at around $7 billion. Aleph Alpha at something between €500 million and $3 billion depending on which outlet you trust. Twenty billion for the combined company is a big step up for both, and it is being underwritten by something that is not a model benchmark. Neither company has a frontier LLM. Neither has a consumer surface anyone outside enterprise procurement could name. What they have, jointly, is a passport that is not American.

That passport is the entire pitch. Cohere has always positioned itself for regulated buyers (defence, energy, healthcare, public sector) and chose enterprise from day one rather than fight ChatGPT for retail attention. Aleph Alpha had German government contracts and, more importantly, a chairman who is friends with people who write procurement specifications. Stitch the two together and you have a credible non-US option for a European Commission that has spent the last eighteen months trying to work out what it actually means to have digital sovereignty when the underlying weights, the underlying chips, and the underlying researchers all come from somewhere else.

The Schwarz piece is the part to watch. Schwarz Group owns Lidl and Kaufland and, less famously, a cloud and data-centre arm that has been quietly scaling for European public-sector buyers. They are building the on-premise hosting infrastructure that German federal procurement will demand. A Cohere model running on Schwarz-operated racks inside a German data centre, sold to a Land government that has been told by Berlin to reduce dependence on American hyperscalers, is a genuinely new product category. It is not a better model. It is a model with a different return address.

I am unsure whether this works. The case for it is that European governments will pay a premium for sovereignty, that defence and healthcare and public-sector workloads are sticky enough to fund a real R&D budget, and that being merely good at enterprise plumbing is a viable place to compete when the frontier labs are pricing themselves at GPT-5.5 levels. The case against is that sovereignty is a soft moat. The moment a European cabinet office decides Claude on AWS Frankfurt is good enough, or decides Gemma 4 on Apache is good enough, the premium evaporates.

There is also a thing that I keep noticing about these consolidation announcements. The word "powerhouse" appears in roughly nine out of ten of them. Mistral and a French national champion. Cohere and Aleph Alpha. The unspoken implication is always that the new entity is now big enough to matter, and the unspoken question is always whether being big enough to matter in your home market is the same thing as being competitive in the actual market for intelligence. So far the answer has been no. The bet here is that Brussels makes it yes.

A 106-year-old working men's club in Droitwich shut its doors earlier this month. The committee cited the usual culprits, rising operating costs, building repairs, debt. The same week, in Cleethorpes, another club went down. Monks Road in Lincoln went in 2018 after a century of trade. The Louth Conservative Working Men's Club rebranded in 2023, dropping the "Conservative", dropping "Working Men's", trying to stay alive as Louth Social Club after membership fell from a thousand to three hundred. The Club and Institute Union itself, the federation that has stitched these places together since 1862, has quietly cut "Working Men" from its own name.

Three-quarters of the country's working men's clubs have closed in the last fifty years. In the 1970s the CIU listed about four and a half thousand affiliated clubs and four million members, a tenth of the adult population. The current figure is around eleven hundred, with some recent counts putting it under a thousand.

It would be tidy to blame the 2007 smoking ban, and people do. The ban hurt, of course. But the decline was already a long-running project, sitting underneath the headline cause. The mines went first, then the mills, then the engineering shops that named the clubs they sponsored. Once the works closed, the membership pool dried up, because the membership pool was the works. A Railwaymen's Club without railwaymen is a strange room. The smoking ban only finished a thing that deindustrialisation had already arranged.

The interiors are what people miss without knowing they miss them, and the interiors are what cannot be photographed back into existence. Flock wallpaper. Formica bar tops curving round to a glass-fronted display of crisps and pork scratchings. A concert stage at one end with a curtained backdrop and a small electric organ. A bingo board screwed to the wall behind the bar, numbered cards in a wooden rack. A committee room with leatherette chairs and minutes in a ringbinder. None of this is heritage in any official sense. There is no listing, no fund, no preservation society. When the building goes, the whole grammar of the room goes with it.

What's properly hauntological about the working men's club is that it was the kind of social institution the internet has not replaced and cannot replace, and yet it has not been mourned as a loss. A bounded community, geographically anchored, with a printed membership card that admitted you to two thousand other rooms exactly like this one in towns you'd never visit. You could walk into a CIU club in Wakefield with a card from Workington and be served. The card was a passport into a country that has now closed its borders.

Like the village hall at Balcombe, these were the institutional architecture of working-class self-organisation, built before the welfare state arrived to do some of the same work, and now outliving the world that made them make sense. The buildings persist longer than their function. A shuttered concert room with the bingo board still on the wall is not a ruin yet, only a room waiting to be turned into flats.

Reverend Henry Solly, who founded the CIU in 1862, was a teetotaler who wanted alternatives to the pub. The members, within three years, voted the alcohol back in. That tells you everything about who the institution actually belonged to. It belonged to them, and they ran it, and now it is closing because there are fewer of them left, and the ones who remain are tired.

Open the firmware of any cellular modem shipping in 2026 and you'll find it answering to a command language designed for a 300-baud modem in 1981. AT+CREG to ask the network for registration status. AT+COPS to pick an operator. AT+CGDCONT to set up a packet data context on a 5G NR carrier. The "AT" prefix stands for ATtention, a convention Dennis Hayes and Dale Heatherington coded into the Smartmodem 300 because they needed a clean way for a host computer to interrupt an in-progress phone call without ambiguity.

Hayes Microcomputer Products filed for Chapter 11 in 1998 and was liquidated the following year. The standard wasn't.

3GPP TS 27.007 is the document that keeps the language alive. The current revision, V18.6.0, was published in May 2024 and runs to several hundred pages of extensions to the Hayes set, all of them prefixed AT, all of them readable as plain text over a serial connection that no modern phone actually exposes to the user. Every 4G LTE and 5G NR chipset, every IoT cellular module, every car telematics box, every emergency satellite modem, all of them speak it. The IoT industry has built a quiet, very-large dependency on a textual interface that originally targeted a serial port on an Apple II.

What's strange about it isn't the survival. Lots of old standards survive. What's strange is that nobody has tried very hard to replace it. There is no AT-Next initiative. No working group is sketching a binary successor. The chipset vendors ship reference firmware with the same command interpreter their predecessors shipped twenty years ago, and the device makers who consume that firmware ship it forward unchanged because rewriting it would buy them nothing and break everything that depends on it.

This is what hauntology in protocol design actually looks like. The thing isn't preserved out of sentiment. It's preserved because the cost of dislodging it always exceeds the cost of one more revision. So the language accretes. +CGDCONT got added for GPRS in the late 1990s. +CEREG for LTE around 2009. +C5GREG for 5G a few years ago. The 1981 prefix never moved.

If you've ever flashed an OpenWrt router, configured a Raspberry Pi LTE hat, or watched a Quectel modem boot, you've seen it. The serial console scrolls past: AT, OK. Then AT+CFUN=1, OK. The handshake grammar of a long-dead manufacturer's late-1970s telephone-line equipment, performed silently inside a device that talks to a satellite.

There's a reasonable argument that this is a problem. The text-based interface is slow, error-prone, and difficult to extend cleanly. Modern modems do expose alternative APIs (QMI, MBIM) for the actual data path, but provisioning, diagnostics, and many control surfaces still go through AT. The replacement layers are bolted on top of, not under, the Hayes layer.

The stranger argument is that this is just what infrastructure does. It calcifies around whichever interface was good enough at the moment of consensus, and it stays calcified for as long as somebody, somewhere, still ships against it. The Smartmodem 300 sold for $279 in 1981. Forty-five years later, several billion new devices a year still type AT before they say anything else.

The bomb scare came on the final day of Milan Fashion Week, October 1995, just as the editors were heading over to Prada. The headquarters got cleared, the police swept the building, the show went on. Whatever Miuccia put on the runway that afternoon was always going to be the news. What she actually put on the runway was a problem.

Avocado green and sludge brown. Murky 1970s tones a critic later said hovered somewhere between shades of slime and mold. Checked kitchen-tablecloth patterns paired with dirty 1950s florals, hand-drawn in a way that looked like the printer had given up halfway. The shoes were clunky T-bar sandals and unorthodoxly low-heeled sliders, the opposite of the strappy follow-me heels the rest of fashion was selling that season. The collection was called Banal Eccentricity. The press, mostly, called it Ugly Chic.

Robin Givhan ran a piece in the Washington Post the following May titled "Ugly is in." Susannah Frankel later wrote, in the AnOther cover story for S/S17, that the term belle laide could have been invented for Miuccia at that moment. Alexander Fury, in a 2014 essay reissued half a dozen times since, called the brown "faecal." All of these are compliments. Read them in sequence and you start to understand what was happening: a designer had walked onto the most commercial week of the fashion year and committed an act of taste sabotage so calculated that the trade press needed two years to catch up.

The cleverness wasn't the ugliness. The cleverness was that the ugliness was made out of the most expensive materials a luxury house could source. Cashmere, silk, the high-tech nylons Prada had been refining since the 1984 backpack. The kitchen tablecloth was hand-embroidered. The avocado wool was woven to the exact gauge a couture house would demand. None of it looked it. That was the point.

Miuccia had inherited the company from her mother in 1978 and spent the eighties quietly building a reputation for understatement. The black nylon backpack of 1984, the gauzy minimalist suits of the early 1990s, none of that prepared anyone for what S/S 1996 actually did. It threw out the playbook of seduction. It said the female silhouette did not have to flatter, and that taste itself was a kind of laziness. Miu Miu, launched three years earlier and named after the family nickname, had been her sketchbook for this. The mainline collection finally said it out loud.

What followed is the part that's hard to remember now because it became the water everyone swims in. The off-key cool Frankel describes, the ironic 1970s palette, the deliberate awkwardness around proportion and footwear, the willingness to make the model look slightly wrong on purpose, all of that is now a default mode for half the labels showing in Milan and Paris. You see a deliberately bad sandal in a 2026 lookbook and the visual grammar comes from this one show.

Miuccia's ugly-chic vocabulary outlasted the supermodel era it interrupted. It outlasted the boom that bought it. It is still, somehow, the cleverest argument a designer has made against beauty in my lifetime, and the only one that the market eventually agreed with.

At Balcombe in West Sussex, set back from the road, there is a hall with a kitchen at one end and a stage at the other. The kitchen was designed to double as a meeting room for the local Women's Institute. The stage was designed for whist drives and amateur dramatics and the reading of parish council minutes. The walls carry murals by Neville Lytton depicting war and peace. The building is called the Victory Hall, and it was paid for by a woman called Lady Denman, who was the first national president of the WI, and it is often described, in the history kept by ACRE, as the first of its kind.

Its kind being the purpose-built English village hall. Which is a thing I did not really understand as a category until I started paying attention to the ones I drove past.

There are thousands of them. Most were built in the decade after the First World War, paid for by the grief of a country that had lost 880,000 men in four years and did not know where to put the feeling. A lot of them are war memorials in the strict sense, with a plaque of names by the door, or sometimes the whole building is the memorial, with the names folded into the act of unlocking it on a Tuesday evening for a yoga class. The Historic England account is clean about this: the halls were a way of converting loss into use.

The machinery that built them is still running, which is the odd part. The Development Commission set up a rural building loan scheme in 1924, and that scheme, passed between departments and renamed and reshaped, is now administered by ACRE on behalf of Defra. A village committee somewhere in Rutland asking about a roof grant in 2026 is asking a question first formalised to help villages bury their dead from the Somme. Nobody on either end of the transaction needs to know this for it to remain true.

What I find myself returning to is the specific shape of the buildings. They are almost always a single volume with a small kitchen bolted on, a stage at one end, and a floor that takes chalk marks well. You can fit a badminton court in the main space, and a jumble sale, and a funeral tea, and a parish council. They were designed to be general-purpose in a way that nothing built now is allowed to be.

The thankful villages get a particular mention in the longer histories of this, the ones where every man came back. Some of those halls are peace memorials rather than war ones, built in a register of quiet amazement that the list of names was empty. You can read the plaque if one exists. There is usually no fanfare.

I walked past one the other week, a low pebbledash building with the initials WI picked out in brick above the door, and the noticeboard had a handwritten poster for a whist drive. Thursday, 7:30. Raffle. No internet address. The building was older than anyone who might attend.

DeepSeek previewed V4 this morning, and the interesting part is not the model. It is the stack underneath it.

The headline numbers are respectable. Two variants: V4 Pro at 1.6 trillion parameters and V4 Flash at 284 billion, both mixture-of-experts, both with a 1 million token context window, up from 128,000 in V3. DeepSeek has named its long-context trick "Hybrid Attention Architecture" and claims world-leading cost efficiency at that window size. Pricing is the familiar undercut-the-frontier gambit. Flash is $0.14 per million input tokens and $0.28 per million output, which sits below Haiku 4.5, Gemini 3.1 Flash, and GPT-5.4 Nano. Pro tops out at $0.145 in and $3.48 out, cheaper than Opus 4.7, GPT-5.5, and Gemini 3.1 Pro. On the company's own benchmarks, V4 Pro is claimed to compete with Claude Sonnet 4.5 on agentic coding and approach Opus 4.5. Benchmarks from the vendor are benchmarks from the vendor, but the last time DeepSeek made a claim like this it was broadly true. None of that is the story.

The story is that Huawei announced, the same morning, that its entire Ascend supernode product line supports V4. Ascend 950 clusters, stitched together with Huawei's Supernode interconnect, running a frontier-class open-weight model. Cambricon pre-announced compatibility within hours. DeepSeek has not disclosed what hardware it used to train V4, and that silence is doing a lot of work, but the inference layer is now explicitly, publicly, and aggressively domestic.

This is what the export controls were meant to prevent, and it is what they arguably accelerated. Nvidia H100s and H200s got harder to ship to China in 2022, harder still in 2023, then harder again. The policy assumption was that constraining the chips would constrain the models. Instead it constrained the supply chain around Huawei, and Beijing poured money into closing the gap. The gap is not closed, Ascend 950 is not an H200, but it is close enough for a 1.6 trillion parameter MoE to run coherently across it, and "close enough" is what matters once a single Chinese lab can ship a competitive open-weight model end-to-end on domestic silicon. That is a political fact, not a technical one. It will be cited in policy papers for the next decade.

The open-weight angle is the other shoe. V4 ships on Hugging Face with weights anyone can download. Americans kept their frontier models closed for commercial reasons that made perfect sense at the time; the result, a year after R1 rattled Silicon Valley, is that the cheapest capable model in every size class is Chinese, the best open-weight model in every size class is Chinese, and now the hardware story is Chinese too. Anthropic and OpenAI have spent the last six months accusing DeepSeek of distilling their outputs to train on, and that may well be true, but it is also a slightly desperate argument to be making at this stage. The ship has sailed. The ship was built somewhere else.

I was skeptical about how much V4 would move the needle given how long the wait had stretched, and how many Chinese competitors, MiniMax included, had shipped credible frontier work in the meantime. The model itself might end up being an incremental update. The hardware announcement is not incremental. It is the first time a Chinese lab has said, on the day of a major release, that its production inference runs on chips Washington cannot block. That changes the shape of the next two years of this race, more than any specific benchmark would.

You can't go back to 1996. Not as a place. Not the streets, not the shop windows lit for an evening that ended thirty years ago, not the bottle of perfume still alive on its top notes, not the future still unopened. That door has shut, and the building it belonged to has been refaced. The angle of light on a particular Tuesday afternoon in late September is not retrievable.

And yet certain years don't behave like other years. They don't recede on schedule. They become weather inside you. They come back as a smell, as a light level, as a particular synth pad arriving in a song you didn't think you remembered, as the blue of a 5pm sky in early October, as the texture of a typeface on a magazine spine, as the feeling of being in a room ten minutes before the news arrived that changed everything.

This is what memory does, and it's also what memory refuses to do. It preserves the emotional truth at extraordinary fidelity, the atmosphere of a year, the specific quality of being alive in it. It destroys the access. Atmosphere, not address. You get back the weather, never the doorway.

A bottle, a record, a photograph, a jacket, a magazine. At the time these were just the furniture of life. They didn't have to mean anything, because you lived inside the world that gave them meaning. Now they've become survivors. That's where the wrongness comes from. The object is still here. The world around it has vanished. It sits in the present carrying an atmosphere that no longer has a home.

It looks innocent, but it's also proof of loss. It says: this happened; you were there; you can't go back. That can feel almost accusing. The feeling is a mismatch. The object promises return and can't deliver. It opens the door a crack, lets the air of that time come through, then refuses entry. Familiar and alien at the same time. Not evil, exactly. Charged. A relic with teeth. If these things worked the way you want them to, you'd be a tourist in your own past. They don't work, and you keep them anyway.

Old perfume is especially powerful, because scent bypasses ordinary distance. A smell doesn't feel like a memory; it feels like the past has walked into the room. But when the perfume has darkened and lost its voice, even the key feels corrupted. The thing that was supposed to restore the past now reads as a damaged message from it. It comforts you by proving the past was real, and wounds you by proving it is unreachable. The objects aren't hostile. They've become haunted.

What I notice is that the rituals around them get more elaborate, not less, the further the year recedes. People build little shrines. A shelf of records arranged in a specific order. A folder of images. A playlist with a specific opening track. A bottle kept sealed in a cupboard with a strip of parafilm around the cap. None of this pretends time has reversed. None of it is naive about that. The shrines aren't a trick to get back; they're a way of keeping faith with the person who was there.

That distinction matters. Nostalgia, the cheap version, wants the place back. What I'm describing knows the place is gone and tends to the trace anyway. It's closer to what Mark Fisher meant when he kept returning to traces of futures that didn't arrive. The corridor stays raised across the field even after the line stops running. The earthwork outlasts its purpose by a factor of three. You can't take the train any more, but the embankment is still arrow-straight, and you can stand on it.

Some years resolve into events and pass through. Other years won't. They settle into the body as a quality of air. 1996 is one of those for me, and I think for a lot of people my age, although I won't pretend it's the same year for everyone. The point isn't the calendar. The point is that the year stops being a date and becomes a temperature. A specific way the afternoon used to feel. A small, irrevocable atmospheric reading.

The shrines are the only way to honour it without lying about it. You don't tell yourself the year is still here. You don't tell yourself you can return. You keep the records arranged. You wear the perfume on a Saturday in spring. You play the synth pad. You let the weather come through.

1996 isn't dead. It's folded into you, the way a season is folded into a bulb. It comes up in its own time, not as the year itself, but as the climate the year left behind.