Vol. I, No. 13
Covering 27 January - 9 February 2025
Monday, February 10, 2025
Late edition · A Relentless publication
All the fortnight that mattered, in technology and in the world, read next to what we were building at the time.
Printed on the internet, which is cheaper than it sounds.
TECHNOLOGY

A cheap Chinese model humbles Silicon Valley, and $600 billion vanishes in a day

DeepSeek, an AI model built by a small Chinese lab for a fraction of the cost, matched the American giants and was given away free. Nvidia lost nearly $600 billion in a single day, and the whole "spend hundreds of billions" thesis wobbled.

The AI industry's confidence took its sharpest blow yet this fortnight, from an unexpected direction. A Chinese startup called DeepSeek released a model, R1, that performed comparably to the best American systems on many tasks, that it claimed to have trained for a tiny fraction of the hundreds of millions or billions the American labs spend, and that it gave away free and open for anyone to run. The market's reaction was violent: on January 27, Nvidia, the company whose expensive chips the entire American AI build-out assumes are indispensable, lost about $589 billion in value in a single day, the largest one-day loss for any company in history, as investors confronted a terrifying possibility, that the frontier of AI might not require the staggering spending, the vast data centres, the endless chips, that the whole $500 billion Stargate thesis rests on.

The panic may prove overdone; DeepSeek's cost claims are disputed, it likely used more and better chips than it admits, and its achievement rests partly on learning from the very American models it undercuts. But the shock was real and the questions it raised will not go away. If a small Chinese lab can approach the frontier for a fraction of the cost, then the assumption underpinning the entire American AI economy, that dominance requires enormous, ever-growing spending on compute, and that this justifies half-trillion-dollar infrastructure bets and trillion-dollar valuations, is suddenly in doubt. And it came from China, despite years of American export controls designed to prevent Chinese labs from reaching the frontier, raising the possibility that the controls have failed, or worse, spurred the efficient, home-grown innovation they were meant to stifle. In a fortnight, the story of AI shifted from American inevitability to genuine contest, and the most expensive assumptions in technology were, for the first time, seriously questioned.

WASHINGTON

The blitz accelerates: an agency dismantled, tariffs launched, and a plan to "own" Gaza

The new administration's assault on the shape of the government accelerated. An effort led by Elon Musk moved to dismantle USAID, the agency that administers American foreign aid, freezing its funding and recalling its staff, effectively shuttering overnight a body that had been a pillar of American soft power for six decades. Trump launched his trade war, imposing then partly pausing steep tariffs on Canada and Mexico and levying new ones on China, sending markets lurching. And in a statement that stunned the world even by the standards of the fortnight, Trump declared on February 4 that the United States would "take over" the Gaza Strip, relocate its Palestinian population, and develop it into a "Riviera of the Middle East," a proposal condemned across the region and the world as a fantasy of ethnic cleansing.

IN BRIEF

Disaster over the Potomac; a game decided

The deadliest American aviation disaster in a generation struck on January 29, when an American Airlines regional jet collided with an Army helicopter over the Potomac River on approach to Washington's national airport, killing all 67 people aboard the two aircraft, in congested airspace whose risks had been flagged for years. And the Philadelphia Eagles won the Super Bowl on February 9, thumping the Kansas City Chiefs and denying them a historic third straight title.

The Column

Five numbers, read slowly

DeepSeek's technical report prices the training of its model at $5.576 million. Nvidia's shareholders priced the news at $589 billion. Between those two figures sit three others, and the five, taken in order, tell the story more carefully than the panic did.

Start with $5.576 million. It is not an analyst's estimate; it is the laboratory's own arithmetic, printed in a technical report published in the last week of December: 2.788 million hours of graphics-processor time, costed at a rental rate of two dollars an hour. The authors fence the figure in the same passage, noting that it covers the final training run alone and excludes everything spent on prior research, failed experiments and the preparation of data. That caveat did not travel. When the lab's reasoning model followed in January, free to download, and its app climbed to the top of the American charts, the December arithmetic, stripped of its own footnote, became the only number anyone discussed, quoted as the full price of admission to frontier AI, which is a claim the report never makes.

The second number is 2,048, the count of chips the run used, and the chip matters more than the count. The H800 is not Nvidia's best processor; it is a legal artefact. When Washington restricted chip exports to China in October 2022, it drew its line partly at interconnect speed, the rate at which one chip talks to another, and Nvidia responded by taking its flagship, cutting that chip-to-chip bandwidth from 900 gigabytes a second to 400, and selling the degraded result into China lawfully. A year later the rules were tightened and the H800 was banned as well, but DeepSeek had bought its fleet while the window stood open. The hardware beneath this model is therefore itself a reading of American law: a processor designed down to the measurements of the regulation, acquired in the interval the regulation left.

The third and fourth numbers describe what was done under that constraint: 671 billion parameters in the model, of which 37 billion are consulted for any given word. The distance between those two figures is the engineering. A model that wakes a twentieth of itself at a time sends far less traffic across the slow links between chips, and the report's other methods point the same direction, arithmetic compressed to eight bits so that less must move, communication scheduled underneath computation so the throttled connections never sit idle. Read whole, the document is a manual for being poor in one specific resource, the interconnect bandwidth American policy had rationed, written by people who had no choice but to become expert in that poverty. Laboratories paying for 900 gigabytes a second never needed such a manual. The lab held to 400 wrote one, and then published it for anyone to use.

The fifth number is $589 billion, the value Nvidia shed on January 27, the heaviest single-session loss any company has ever taken. The market, in effect, accepted the first number at face value and concluded that the frontier no longer requires the spending that justifies the chipmaker's price. As a piece of accounting that overshoots badly: the $5.576 million excludes the research that made the run possible, the hedge fund behind the laboratory disclosed owning ten thousand of Nvidia's previous-generation chips years ago, and sceptical analysts put the true cost of the whole programme in the hundreds of millions. But correcting the arithmetic does not retire the problem, because the methods are published, reproducible by any team whose bandwidth is capped, and every team the controls were aimed at is a team whose bandwidth is capped. The rule set its ceiling at 400 gigabytes a second, and what the report documents, figure by figure, is the engineering that grew to fit beneath it.

Field Notes
A Relentless build, told plainly

Making the systems that will not talk to each other talk to each other

A company is not one system but dozens, and each holds a different version of the same customer. The most valuable work I do is not building any of them; it is making them agree. No client is named.

The single most valuable thing I have repeatedly done, across many years and many organisations, is also the one that sounds the least impressive when I describe it: I made systems that could not talk to each other talk to each other. A company does not run on one system. It runs on many, a sales system, a finance system, a system that came with an acquisition, a system a department bought without telling anyone, a decades-old system that outlasted everyone who understood it and that no one will risk replacing, and each of them holds a piece of the truth about the business, and none of them agrees with the others, because they were never built to. The customer is one person, but they exist, differently, in eight systems, with eight slightly different addresses and balances and histories, and the business, made of these disconnected pieces, quite literally does not know its own customer, because the knowledge is scattered across systems that do not speak.

Integration is the work of fixing that, and it is far harder than it sounds, because the difficulty is never technical, or not mainly. The technical part, moving data from one system to another, is a solved problem with mature tools. The hard part is that the systems disagree about what things mean. One system thinks a "customer" is an individual; another thinks it is a household; a third thinks it is an account number. One counts revenue when the deal is signed; another when the money arrives. To integrate them, you cannot just move the data; you have to reconcile the meanings, decide, for the whole organisation, what a customer is and when revenue counts, and then make every system defer to that shared truth. That is not an engineering problem. It is a negotiation, conducted in software, between departments that have spent years quietly meaning different things by the same words and never had to reconcile because their systems kept them politely apart.

I raise this in the fortnight of DeepSeek because there is a connection, and it is about where the real value in technology actually lives. Everyone is transfixed by the models, the frontier, the raw capability, and that is real. But most of the value most businesses will ever get from technology is not at the dazzling frontier; it is in the connective middle, in getting their own systems to agree with each other so that the business can see itself clearly and act on the truth rather than on eight contradictory versions of it. A company with a brilliant AI model bolted onto a landscape of systems that disagree about who its customers are has built a genius on top of a swamp. The integration, the patient reconciliation of what the systems mean, is the foundation that makes everything above it worth having, and it is the work that gets deferred, because it is difficult and thankless and shows nothing on a screen, right up until the business discovers it has been making decisions for years on data that never agreed with itself. Make the systems talk. Make them agree what the words mean. The connections, not the nodes, are where a business's value and its pain actually live; and underneath almost everything, the connections are the work.

The Back Page

The two readings of a near-miss

The corridor into Washington's national airport had produced near-misses for years before this fortnight's collision, and the record of them is the strangest kind of document a dangerous system generates, because it carries two opposite readings at once. To the pilot who files the report, the event is proof of danger: the margin was seconds, the geometry was blind, the helicopter was where the jet needed to be. To the system that receives the report, the same event reads as proof of function: the margin held, the crews reacted, nothing was bent, the safeguards did their job. Both readings are true of the same ninety seconds, and over years the second one quietly wins, because every near-miss that ends safely lands in the file marked works. The sociologist Diane Vaughan, studying how NASA talked itself into launching Challenger, gave this drift a name, the normalisation of deviance: each escape makes the next exposure feel more routine, until the abnormal has been redefined, in good faith and in writing, as acceptable. Nothing about the sky over the Potomac changed on the night of January 29. What ran out was the quantity the file had been spending all along, the margin itself, which had been booked as evidence of safety every time it saved someone, and which on this night had nothing left over for the 67 people it was supposed to be protecting.

Corrections & Amplifications
The Fortnight's Fable

The Cartwright's Wheel

One cartwright made wheels that lasted thirty years; a rival made wheels that lasted three and cost a tenth as much, and the town bought the cheap ones gladly, and bought them again, and again. After thirty years they had paid for wheels three times over and had spent as many afternoons stranded at the roadside as they had saved coins at the fair, though only the coins had ever been counted.

The cheap wheel is the dearest in the end, but it is paid for in small coins on fair days and in whole afternoons stranded on foul ones, and a town remembers only the coins.

The Funnies
FORECAST REALITY
The forecast and the reality, drawn to scale.
The forecast has been revised to match. It is now extremely accurate.
The Oracle

The fortnight read across the heavens, Western and Eastern. Find your sign; the Oracle is never wrong, and never responsible.

The Western Stars
Aries21 Mar to 19 Apr

Your bold new plan is excellent and should be shown to absolutely no one who has tried it before.

Taurus20 Apr to 20 May

Comfort beckons. So does the quarterly review; attend the first, dread the second.

Gemini21 May to 20 Jun

You will say two contradictory true things and be quoted on the wrong one.

Cancer21 Jun to 22 Jul

Protect your peace. It is stored, unwisely, in a single unanswered email.

Leo23 Jul to 22 Aug

You will shine, briefly, before someone reboots the room.

Virgo23 Aug to 22 Sep

You will reconcile two spreadsheets that disagreed by a single cent, and feel, for one clean hour, that the world is ruled by sense.

Libra23 Sep to 22 Oct

Asked where you fancy going, you return the question so gracefully it bounces between you for days.

Scorpio23 Oct to 21 Nov

You will maintain a grudge so immaculately preserved it could be loaned to a museum.

Sagittarius22 Nov to 21 Dec

You will answer 'does this suit me' with total honesty and spend the rest of the fortnight making amends.

Capricorn22 Dec to 19 Jan

You measure the day purely in outputs and go to bed convinced you have, somehow, underperformed at sleeping.

Aquarius20 Jan to 18 Feb

A feeling approaches; you observe it anthropologically, take a few notes, and politely decline to participate.

Pisces19 Feb to 20 Mar

You will mistake a daydream for a plan and seem gently surprised when the universe declines to deliver it.

The Chinese Zodiac
Rat1996 · 2008 · 2020

The pantry you have been quietly stockpiling will outlast three of your colleagues and one of your friendships.

Ox1997 · 2009 · 2021

A colleague tries to hurry you; the colleague tires first.

Tiger1998 · 2010 · 2022

A dramatic entrance is undercut by the fact that nobody had been told you were coming.

Rabbit1999 · 2011 · 2023

Your softness is mistaken for weakness by a person who will shortly need a favour.

Dragon2000 · 2012 · 2024

Pride goes before a staircase you insist on descending slowly, so all may see.

Snake2001 · 2013 · 2025

Coiled patience outlasts a frantic colleague, who mistakes your stillness for agreement.

Horse2002 · 2014 · 2026

You cannot be reined in, which your manager has stopped trying and started documenting.

Goat2003 · 2015 · 2027

You will say yes to be agreeable and spend the week quietly discovering what, exactly, you agreed to.

Monkey1992 · 2004 · 2016

A trick lands beautifully; the audience, being your employer, enjoys it rather less than you do.

Rooster1993 · 2005 · 2017

You will be right, punctually, about something nobody wished to discuss at that particular hour.

Dog1994 · 2006 · 2018

Someone you love does something unwise, and you appoint yourself, unpaid, to pace about it.

Pig1995 · 2007 · 2019

Someone offers you rather more than you need; you accept it, warmly, on principle.

The Vedic Rashis
MeshaMangal · 14 Apr

Mangal rides hot in your house; the fire you light to warm the room is the same fire that burns it.

VrishabhaShukra · 15 May

Wealth gathers slowly in your corner, like dust that happens to be gold. Do not mistake stillness for poverty.

MithunaBudh · 15 Jun

Speech is seed. What you scatter lightly now, the wheel returns to you grown, rooted, and hard to uproot.

KarkaChandra · 17 Jul

The mind's waters are muddied now. The sages counsel stillness: let the silt settle before you look for your face.

SimhaSurya · 17 Aug

A radiance you spent on applause could have warmed one cold hand instead. Surya notices both, and weighs them.

KanyaBudh · 17 Sep

Detail is devotion when offered, tyranny when imposed. Mercury asks which of the two you are practising today.

TulaShukra · 18 Oct

Balance is a verb the planet conjugates daily. You will not arrive at it; you will keep choosing it.

VrishchikaMangal · 17 Nov

Mangal buries your fire deep, where it burns longer. A secret kept this fortnight becomes a kind of power.

DhanuGuru · 16 Dec

You aim your arrow at the horizon and call it a plan. Jupiter admires the optimism, funds half of it.

MakaraShani · 15 Jan

A duty you resent is quietly becoming your foundation. Shani builds with stones you would have discarded.

KumbhaShani · 13 Feb

Detachment is not coldness. Saturn loves at a distance, the way a star warms a world it never touches.

MeenaGuru · 15 Mar

The tide of the unseen pulls you inward. Do not fight it; Guru charts the depths you fear.

A heavy night. The front page would not settle, the way big news never does until it is too late to change it.Set after hours by the Night Compositor