Mikhail Parakhin: Shopify's Unlimited Tokens and Barbell AI Strategy

Shopify CTO Mikhail Parakhin explains why the company gives every engineer unlimited AI tokens, always uses the largest model, and built a digital twin of its merchants. In this 12-minute conversation on Anthropic’s Claude channel, he talks with Boris Cherny, the creator of Claude Code, about the difference between using AI to do old work faster and using it to do work that was never possible, how Shopify measures engineering productivity, what AI changed about managing people, and the “barbell” token strategy he recommends to other CTOs.

TLDW

Parakhin argues that the real payoff from large language models is raising the ceiling (solving problems that no amount of people or time could solve before), not raising the floor (automating drudgery). Shopify’s proof is a prediction system that treats each merchant as a sequence of actions, builds a digital twin of the business, and tests interventions such as a loan or an ad campaign on the twin before offering them for real. His policy at Shopify is unlimited tokens and the heaviest models for everyone building software, with circuit breakers for runaway processes. His advice to other technical leaders is a barbell: spend freely on the largest model for development, testing and research, then save money in production with smaller, fine-tuned models.

Thoughts

Parakhin’s barbell is an accusation as much as a strategy. He says he has repeatedly seen companies ration coding tokens for their engineers while running models that are larger than they need in production, which is the spending pattern exactly backwards. The economics support him. Development tokens are a small line next to engineering salaries and they buy the decisions that everything else depends on, while production inference scales with traffic and compounds forever. Two caveats belong next to the advice. This is a conversation on the channel of the company that sells the tokens, so “always use the largest model” is advice the host has no reason to challenge, and nobody in the room does. And the cheap end of the barbell is not free: fine-tuning a right-sized production model takes GPU capacity and machine learning talent, which Shopify has and many companies do not.

The digital twin is the most interesting thing described here and the least examined. Predicting a merchant’s next action from its history is a natural fit for a sequence model. Asking what would happen if Shopify offered that merchant a loan is a different kind of question, because it is causal, and a model trained on what merchants did by themselves does not automatically know what happens when someone intervenes. Parakhin says Shopify acts on these predictions in production and that the system has a direct impact on the bottom line. He gives no numbers, and the conversation does not ask how the counterfactuals are validated. Shopify does have an unusual advantage, since it can run the experiment for real across a very large merchant base and feed the results back. That feedback loop, more than the model, may be the thing competitors cannot copy.

His sharpest warning is about what gets lost when building becomes cheap. A model tries to please, he says, so the answer is often already sitting in the question, and it will hand you exactly what you asked for. Before, a half-formed idea had to survive a discussion before anyone built it. Now it can be built before lunch. That makes the old friction look like a quality filter that companies got for free and are now removing. His conclusion is that technical judgment matters more for leaders, not less, which is a useful correction to the idea that AI makes everyone a builder.

The weakest answer is the one about return on investment. Cherny asks how a CTO should justify ceiling-raising work to a CEO when it cannot be measured the way saved engineering hours can. Parakhin describes in detail how well Shopify measures the floor, then answers the ceiling question with a default: use the maximum until someone shows evidence against it, and make sure the people deciding actually use the product. That may be the honest answer. It is also a position that only works at a company whose CEO is already convinced, and it will not help a CTO who is facing a skeptical finance team.

Key Takeaways

Parakhin says he used Claude to solve a math problem he had been stuck on for seven years, involving the Wasserstein loss function, which earlier models could not crack.
He separates two kinds of AI value. Raising the floor is doing existing work faster. Raising the ceiling is doing what was impossible before, and he considers it far more important because diligence or headcount could always raise the floor.
The best problems today need both sides. He cannot solve them without the model, and the model cannot solve them without him. He compares it to centaur chess, where a human and an engine together beat either alone.
Shopify measures floor-raising beyond counting pull requests. It estimates the complexity of whole projects, then tracks how many projects teams finish and how fast, normalized for that complexity.
Shopify gives everyone unlimited tokens and runs the heaviest models it can find. The reasoning is that without trying the largest model you never learn what you are missing.
His rule for the CEO conversation is to make sure every decision maker is personally using the models and the product, then keep pushing the ceiling until the returns are obvious.
Shopify’s merchant prediction system represents a company as a sequence of actions (open a credit line, ship a product, click an ad, take a loan), the same way a language model treats text as a sequence of letters.
That sequence becomes a digital twin. Shopify tests counterfactuals on it, such as an advertising campaign, shipping one day faster or a loan, then makes the best offer to the real merchant in production.
Parakhin believes the system was not humanly feasible before, because nobody would have assembled the talent or started collecting the data. His definition of raising the ceiling is doing something that was not even in the consideration set.
He changed his mind about AI and management. He built a system that analyzes what is happening across his organization and warns him when a project looks likely to slip or when people are unhappy.
Because models try to please, it matters more than ever to ask for what is actually needed. Managing engineers who supervise models feels to him like becoming a second-level manager.
The barbell strategy: largest model for coding, development, testing and research, with circuit breakers for runaway processes, and cheaper, often fine-tuned models in production.

Chapters

0:35 Raising the Ceiling: A Seven-Year Problem Solved With Claude

Asked where Claude has surprised him, Parakhin says he finally found a solution to a problem he had worked on for seven years, involving the Wasserstein loss function. He says the point is not that models remove drudgery so people can do more interesting work. It is that they make possible things that no amount of time or help could have achieved, even with thousands of the best mathematicians. The best problems require back and forth between person and model, which he compares to a centaur in chess. Cherny sums it up as raising the ceiling as well as the floor, and Parakhin agrees that the ceiling matters much more.

2:07 Measuring AI ROI and Engineering Productivity

Cherny points out that floor-raising is easy to quantify, since work that engineers would have done is now done by Claude, and asks how to value work that was out of reach before. Parakhin answers first on the floor, where he claims Shopify is probably the best in the world at measurement. Like others, it estimates the number and complexity of pull requests. It also estimates the complexity of entire projects in its internal project system, so it can see that teams are finishing more projects, faster, after normalizing for difficulty. He says this allows very precise productivity percentages.

3:32 Why Shopify Gives Every Engineer Unlimited Tokens

On how to have the spending conversation with a CEO, Parakhin’s rule is to always use the largest model, because it raises the ceiling the most and because you cannot know what you are missing unless you compare. His default is to throw the maximum at a problem until he sees evidence that it is not needed. That is why Shopify has unlimited tokens for everybody, runs the heaviest models it can find, and invests heavily in GPU inference, training and optimization. Everyone making the decisions has to be using the model and the product themselves, he adds. After that, whether it works becomes obvious.

4:42 Building a Digital Twin to Predict Merchant Growth

The project Parakhin is most proud of is a prediction system built on one idea: a company can be represented as a sequence of actions, the way text is a sequence of letters. Opening a credit line, shipping a product, clicking an ad and accepting credit cards are the large events, but everything a company and its employees do can be turned into that sequence. The result is a digital twin that Shopify can experiment on instead of the real business. It asks what would happen with an ad campaign, faster shipping or a loan, finds the sequence of interactions that maximizes the merchant’s chance of growth, then makes that offer or sends that advice in production.

6:45 Why It Could Not Have Been Built Before

Cherny asks how long the twin would have taken in the past. Parakhin rejects the premise and says he does not think it was possible. A company would have needed the concentration of talent, the premonition that the thing was worth building, and the patience to start collecting the data. His honest answer is that Shopify simply would not have done it. That is his definition of raising the ceiling: starting on something that was previously not even considered.

7:28 AI for Managing Engineering Teams

The thing Parakhin changed his mind about is management. He believed language models could not help him run a team better, and now says he could not have been more wrong. He built a system that analyzes what is happening and tells him proactively when a project looks like it will slip its deadline. It lets him catch problems before they happen and notice when people are unhappy, and he says the whole organization functions better for it.

8:10 How the CTO Role Changed: Technical Judgment Matters More

Parakhin says the technical side of his job is now more important. A model is good at picking up signals and tries to please, so it will build whatever the question implies. Putting the first thing that comes to mind into words can produce worse outcomes than before, because there is no longer a discussion standing between an idea and its implementation. Managing engineers who supervise models, and not engineers who type, is like becoming a second-level manager. The new questions are how to dispatch work across models, how to keep track of what they are doing and how to avoid conflicts between them.

9:41 Advice for CTOs: The Barbell Token Strategy

His main advice is to try not to limit tokens, while keeping circuit breakers for runaway processes that start burning them. For anything in coding, development, testing and research, use the largest model and budget for the cost. For production, fine-tune the right-sized model and research it hard to balance cost, throughput and quality. Save money in production and invest it on the development side. He calls the move to Claude Fable 5 a much bigger step than all the previous ones, and says he will outcompete anyone who does the opposite. He closes with a joke about Linear A, the undeciphered Minoan script he had planned to tackle in retirement, until Cherny posted that someone had done it with Claude.

Notable Quotes

“I cannot solve it without the model, but the model cannot solve it without me either.”
Mikhail Parakhin, on the best problems he works on with Claude

“So you’re raising the floor. But actually, don’t forget, you can also raise the ceiling.”
Boris Cherny, summing up Parakhin’s argument

“You can create a digital twin of the company, and then you can start doing things to it instead of the real company and then see what happens.”
Mikhail Parakhin, on Shopify’s merchant prediction system

“I think you start doing something that previously was not even in consideration set.”
Mikhail Parakhin, defining what raising the ceiling means

“The LLM is very good at picking up signals and it tries to please. And so whatever you ask, the answer is very often in the question.”
Mikhail Parakhin, on why technical judgment matters more for leaders now

“It might not look like there’s big difference between big models and small models. There’s a world of difference and you want it.”
Mikhail Parakhin, on why he does not limit tokens

“Boris tweeted that somebody did it using Claude, and I’m like, damn it, that was my whole spiel.”
Mikhail Parakhin, on losing his retirement project of deciphering Linear A

Watch the full conversation between Mikhail Parakhin and Boris Cherny here.

Mikhail Parakhin: Shopify’s Unlimited Tokens and Barbell AI Strategy