In October 2021, inspectors from the US Food and Drug Administration went through years of operating records at a factory in Brussels. On one filling line, they found, the air-filtration system had failed again and again between 2017 and 2021, and sterility had been “compromised” in the area where drugs were made. Staff had logged the failures as “minor.”
The plant belonged to Catalent, a contract manufacturer that fills syringes for drug companies. After the inspection it shut down. One of the things it filled was the syringe inside the pens for Wegovy, Novo Nordisk’s new weight-loss drug. That November, Novo had said US demand had “significantly exceeded our expectations.”
On December 17, Novo told investors that a contract manufacturer filling syringes for Wegovy pens for the US market had stopped deliveries. It no longer expected to meet US demand in the first half of 2022, and “few new patients are expected to be able to initiate treatment.”
Novo’s announcement pointed to the filling, not to a shortage of semaglutide, the medicine itself. Its annual filing a few weeks later said no raw material shortage was expected to significantly affect supply to any major market. What had stopped was one of the last steps before a dose leaves the factory. A Novo spokesperson later told Reuters that “it would have been good to have more capacity and thus, more than one facility for supplies.” The case follows the next three years.
One word
The TOC Institute tells a story about Eliyahu Goldratt, the physicist behind the theory of constraints. A reporter challenged him to sum up the theory in a single sentence. Goldratt said he could do it in one word: focus.
The five focusing steps are what that word means in practice:
- Identify the constraint, the step that limits what the whole system can produce.
- Exploit it: get everything you can out of it before buying more.
- Subordinate everything else to it, so the rest of the system works at its pace and keeps it fed.
- Elevate it: now spend money to expand it.
- Don’t let inertia become the constraint. Once it is no longer the weakest link, something else is. Go back to step one.
The order matters. The institute warns that people tend to skip the first three steps and go straight to the fourth, the expensive one.
Read against the steps
Novo didn’t describe its response in these terms, and its moves didn’t come in this order. Some of them still map onto the steps.
Its first announcement pointed to filling as the step that had stopped. Novo’s priority, it said, was patients who had already started treatment. That was a way of deciding who got a limited supply, which is close to exploiting a constraint but not the same thing: nothing in the record shows it got more pens out of the filling lines. In May 2023, with demand still ahead of supply, it reduced US supply of the lower dose strengths “to safeguard continuity of care.” Those are the starter doses new patients begin on, and Reuters reported the cut was by half.
The rest of the company was paced to the filling step. Wegovy had been approved in Britain and the EU, but “due to the current supply constraints,” the 2021 annual filing said, only a limited launch outside the US was expected in the second half of 2022.
It also elevated. A second contract manufacturer began producing Wegovy in April 2023, a month before the cut to starter doses. In February 2024 Novo agreed to take over three Catalent plants that specialize in sterile filling, for $11 billion up front, expecting them to raise its filling capacity gradually from 2026. One of them was the plant in Brussels.
In January 2024, Novo began gradually increasing supply of the lower doses again. A year later, the FDA declared the semaglutide shortage resolved. That is where the fifth step would start. A rule built for a shortage, like rationing starter doses, is the kind of habit it warns about: when the constraint eases, the rules made for it have to be checked again.
The pizza oven
Imagine a small pizza restaurant with one wood-fired oven that fits four pizzas. On a Friday night the oven is the constraint. Exploit it: the next pizzas are topped and waiting before the last ones come out, so the oven is never half empty. Subordinate to it: the front of house stops promising twenty minutes when the oven is forty minutes behind, and nobody preps forty bases that will sit and dry out. Only then elevate: a second oven. And once it arrives, the owner might find the dough station is now the limit, and Friday works differently again.
When the constraint is a rate limit
When a team runs a lot of work through Claude’s API, the constraint can show up as an error message. Anthropic’s API returns a 429 error that says which rate limit was exceeded: requests, input tokens or output tokens per minute.
The steps still apply. To exploit an input-token limit, Anthropic’s documentation points to prompt caching. For most Claude models, only uncached input tokens count toward that limit. Its example: with a limit of 2 million input tokens a minute and an 80% cache hit rate, you could process 10 million input tokens a minute. To subordinate, smooth the flow. The docs warn that a limit of 60 requests a minute might be enforced as one a second, so short bursts can trigger errors. Work that can wait can go through the Message Batches API, which has its own limits. Only then ask for higher limits.
And then look again. Once the model can do more, the constraint may be the person who has to read what it produced.