What was inside the hour

Z.ai published the weights of GLM-5.3-Flash on Hugging Face on 26 August 2026 under an MIT licence, the most permissive in common use ^[1]. The model declares three hundred and twenty billion total parameters, activates eighteen billion per token, and is natively multimodal ^[1]. Its list price is fifteen cents per million input tokens and fifty cents per million output, with cached input at three cents ^[1]. The licence matters here as much as the tariff, because whoever downloads those weights adapts them to their own organisation, runs them on their own infrastructure and keeps whatever comes out, with no royalty and no permission.

Running those weights in-house is not free either, and that is worth saying before it starts to sound like magic. The FP8 version takes up some three hundred and six gigabytes, wants Hopper-generation cards or newer, and demands people able to keep the deployment standing ^[1]. What the licence removes is not the cost but the permission, because nobody can cut off your access, change your terms or raise your rate halfway through a contract.

The larger phenomenon of which that release is today the best available example is not about models. The unit professional work has been billed in for a century — the hour — is being retired from the market, and what retires it is the collapse of the input that justified its price. Nobody discovered that unit's flaw this week. William Rehnquist, then Chief Justice of the United States, told the graduating class of the Catholic University law school in 1996 that the billable hour rewards inefficiency ^[2]. His image was crueller than his diagnosis, because a firm that prices itself by the hour differs little from the office-supply vendor who counts the pencils delivered and sends a bill for that amount ^[2]. Thirty years later somebody put a price on the warning. The hour never charged for the result of the work but for the time of the person doing it, and that substitute worked while expert time was the scarce input of the operation. That assumption is the one that has just broken.

Whoever sells cheap intelligence bleeds too

Z.ai is not an academic experiment. The company booked 953.89 million yuan in the first half of 2026, some one hundred and forty-two million dollars, a rise of close to 400 % ^[3]. Its open platform and API business multiplied by twenty-seven and now accounts for 86.5 % of revenue, where a year earlier it was 15.2 % ^[3]. Net losses narrowed to 2.07 billion yuan ^[3]. Bloomberg still ran the headline that sales came in below estimates, punished by the Chinese price war ^[3]. Selling intelligence at fifteen cents leaves little margin. Whoever buys at that price takes away for almost nothing the input that until recently justified a professional rate, and that asymmetry between the seller and the user orders everything that follows.

The large buyers stopped paying by the hour

Technology consulting will move four hundred and twenty billion dollars in 2026, 8 % more than last year, and two hundred and thirty-six billion of that figure is implementation work, which is the exposed segment ^[4]. Its clients have already dug in. Greg Meyers, of Bristol Myers Squibb, says his technology support costs are collapsing, pushes his advisers towards fixed-price contracts or performance-related fees, and warns that it has never been easier to take the work in-house ^[4]. UniCredit cut its consulting spending by 24 % in the first half of 2026 and Société Générale by 9 % ^[4]. Commerzbank is aiming at five hundred million euros of annual savings by 2030, and its head of artificial intelligence explains that he now settles in days the analyses that used to take consultants months ^[4]. Jochen Kamp, of Bayer, sums up the buyer's mood in three words, because he expects that “fewer and fewer” traditional consultants will be needed as agents take over coding and testing ^[4].

The software vendor SAP holds that its changes can halve the cost of external consultants, and its chief financial officer states that AI will massively displace that work ^[4]. Aiman Ezzat, who runs Capgemini, answered by calling the figure “ambitious” ^[4]. The markets have already chosen whom to believe: Capgemini is down 31 % so far in 2026 and Accenture 27 % ^[4]. Only a third of clients rate their transformation projects as entirely successful, and the share planning to use the four big audit firms more fell from 80 % to 55 % in a year ^[4].

The trade that invested most saw its hours rise

American law firms closed 2025 with their best year of growth since the financial crisis, profits up 13 % and worked rates up 7.3 %, more than double inflation ^[5]. They bought technology as never before and expanded headcount at the same time ^[5]. Lawyers at the hundred largest firms now cross a thousand dollars an hour while the market average runs around six hundred, and past a certain distance the corporate client starts making different decisions ^[5]. The Thomson Reuters report that gathers those numbers describes a “structural business model conflict”: the industry sits trapped between a transformative technology and an outdated billing structure that “may no longer reflect the value delivered” ^[5]. The figure holding up the diagnosis is a single one. 90 % of legal dollars still flows through hourly billing arrangements ^[5].

It pays to be precise about what that number proves and what it does not. Aggregate hours did not fall, and that does not demonstrate that each task still takes as long. It demonstrates that the billing unit absorbed the improvement without changing its name. A practice that speeds up its work and goes on billing by the hour has only two exits, which are charging less or going on writing the same hours. Thomson Reuters named the tension the two-thousand-dollar-hour problem ^[6].

Where the unit has already moved

WPP, the advertising group that owns Ogilvy, is taking a quarter of its net sales out of the time-and-materials regime ^[7]. Its chief financial officer, Joanne Wilson, told analysts that a commercial model tied to client outcomes would let the company “move away from time and materials and decouple revenue from headcount” ^[7]. That sentence describes the whole movement, because the hour tied revenue to the number of people and that link is exactly the one cheap intelligence breaks.

Globant offers the regional case, and it is the most instructive because it shows both halves at once. The company reported quarterly revenue of 614.4 million dollars on 13 August, flat against the prior year. Its adjusted operating margin came in at 13.2 % and missed the guidance it had set itself ^[8]. It cut the annual forecast and its share price fell 17 % the following day ^[8][9]. Management attributed the cut to new markets, to the slowdown in travel and hospitality, and to longer decision cycles in North America; not to artificial intelligence ^[8].

The other half sits in the same presentation. AI Pods — teams that combine engineers with agents and are billed on outcome — took Glob.AI's recurring revenue to 52.8 million dollars, 61 % above the previous quarter, and the year-end target rose to at least one hundred and ten million ^[8]. The company describes the new unit in its own words, because on its platform the client “pays on the output or consumption they receive rather than on the hours behind it” ^[8]. The decisive number is another one. The gross margin on those contracts runs some ten percentage points above traditional delivery, because fixed pricing lets the firm keep the productivity gains it generates itself ^[8]. Those ten points are the measurement of what the old unit costs.

The mechanism fits in a single sentence: the hour charges for the input, so that every productivity gain travels to the client as a shorter invoice, and the firm that uses the tool best ends up being the one that loses the most revenue. The ten margin points Globant reports on its outcome-based contracts are therefore not an operating efficiency but the portion of the gain the hour had been giving away ^[8].

Holding on to the old unit has defenders with concrete motives, and it pays to name them rather than put it all down to inertia. The partner's profit comes from leverage, from the spread between what a junior costs and what a junior is billed at, and charging on outcome dismantles that pyramid. The people who must approve the change are exactly the ones the change demotes. Junior work was also the quarry the seniors were cut from, so automating it leaves the firm without a bench ten years out. Charging on outcome forces the firm to absorb an execution risk that used to travel to the client, and that demands capital and insurance the mid-sized firm does not have. Buying is built in hours too: a tender asks for a rate, not for an outcome.

Latin America did not sell hours. It sold cheap hours, which is a far more fragile position, because the discount only makes sense while the discounted thing exists. Argentine exports of knowledge-based services reached 10.085 billion dollars between April 2025 and March 2026, 11.7 % more than the previous period, and are now the country's third-largest generator of foreign currency behind agriculture and energy ^[10]. 72 % of the sector's companies name cost competitiveness as their determining factor ^[10]. The region's modern services concentrate in Brazil with a third of the total, Mexico with 17 % and Argentina with 10 % ^[11]. All of that supply is a discount on the unit being retired, and no discount holds up when what disappears is the unit.

The counterpoint has to be conceded whole, and it has three legs. This looks more like a transition than an extinction, and Globant's own numbers show it, with the new business growing 61 % in a quarter while the legacy flattens ^[8]. Consulting never sold only knowledge either: it sells an outside signature that authorises an internal decision, and a model cannot be blamed, nor fired, nor seated in front of a board. That 90 % of legal dollars still travelling by the hour is not only inertia, but also a buyer paying for cover and knowing it ^[5]. The third leg is that somebody is paying for cheap intelligence, and for now the seller is: Z.ai lost 2.07 billion yuan in the same half in which it quintupled its revenue ^[3].

THE PLAY

1. Move a single service line before you are asked to. Pick the one where the tool has already opened the widest gap between what you take and what you bill, put a fixed price on it and keep the difference. The ten points exist, and today you hand them to the client as a shorter invoice.
2. The asset is not the model: it is the context. An MIT licence lets you adapt the weights with an organisation's documents, tickets and processes, and that artefact stays where it was trained. Write into the contract who keeps it, because for forty years the client paid for the discovery and the supplier walked off with it.
3. Sell the conversion to your own trade. The region concentrates thousands of firms that bill by the hour and buyers who still do not know how to buy outcomes. Drafting that contract, measuring the outcome and setting the price is a service that does not exist today, and the market is next door to your office.