Good morning! Welcome to the Canadian AI Newsletter, a weekly rundown for founders, operators and investors.
Was this forwarded to you? Subscribe below.
I am Raif Barbaros, Partner at Mistral Venture Partners. Views are my own.
This is a different kind of issue. Every quarter, we send our LPs a report on each fund, with a market update, company updates, fund financials, and a section I write on where we see AI heading. This quarter I took on a question I keep hearing from founders: if capable models keep getting cheaper and closer together, where does the value go? Below is that section, unchanged apart from light formatting. It's global rather than Canadian in scope, but what it describes resonates here too. The regular Tuesday issue returns next week.
From our latest quarterly report:
A year ago, we wrote that open-source AI was tracking the same path as Linux and PostgreSQL: the model layer would commoditize, and value would move up the stack. One year later, that call has mostly played out and faster than we expected.
DeepSeek released V4 in April with fully open weights, priced at roughly a tenth of comparable Western APIs and with near-frontier coding performance. Alibaba’s Qwen models have passed one billion cumulative downloads. Yet the Western “Red Hat of AI” we hoped to see has not emerged, and leadership in open models is even more concentrated in China than a year ago.
The bigger surprise is that the closed frontier is commoditizing too. In July, five labs released flagship models within two weeks, with results close enough that several clear the bar for many enterprise uses. It reminds us of digital cameras in the 2000s: once every camera had enough megapixels, buyers stopped comparing sensors and started comparing lenses and software. Enterprises are approaching that point with AI. Buyers used to ask which model is best. Now they ask which models are good enough, and which one offers the best price, speed, reliability, and control.
Why, then, do the model companies keep growing so quickly? Falling prices can be offset by expanding consumption, and the leading labs are moving up the stack. The best available estimates suggest that more than half of Anthropic’s revenue comes from coding, while roughly 70% of OpenAI’s comes from ChatGPT subscriptions. Neither discloses the exact mix, but Anthropic seems to have been propelled largely by a single enterprise workload (coding) and OpenAI by a general-purpose consumer product. Neither has built its growth across a broad set of enterprise workflows yet. That market is still early.
Uber recently showed what a sophisticated enterprise buyer looks like. Following its Q2 earnings, CTO Praveen Neppalli Naga wrote that Uber had more than quadrupled the number of engineers using frontier AI tools since January, even as its cost per token declined. Engineers see their costs in real time, while Uber tests proprietary and open-weight models and uses the least expensive option that meets each task’s requirements. He described it as the end of “tokenmaxxing,” when consuming more tokens was itself treated as progress.
This is enabled by the harness: the software between a company’s workflows and the models that handles routing, caching, evaluations, and monitoring. A good harness turns switching models into an operational decision rather than a rebuild. Companies can then route work to the least expensive model that clears the quality bar. CIOs have always hated vendor lock-in, so this hits home.
None of this means every model is identical. But capable intelligence is getting cheaper, more abundant, and easier to swap out, and enterprises are starting to buy accordingly. That is good news for nearly everyone building above the model layer, and the broad enterprise market is still up for grabs. As enterprises optimize their spend at the model layer, our going-forward focus stays on the picks/shovels and application layers where we believe the next phase of value capture will take place.
See you Tuesday.
-Raif