AI Newsround – TSMC & China

TSMC: Good but in line.

• TSMC reported excellent sales for the month of June, but this is well within the range of expectations, meaning that a blow-out on Thursday (Q2 26 results) is now increasingly unlikely.
• Revenue for Q2 26 was TWD1.7bn ($39.6bn), representing strong growth, but this was already well known as revenue came in broadly in line with the consensus estimate.
• These numbers underscore the fact that the AI boom is still on, but ask the question of to what degree this is already baked into forecasts and valuation?
• To me, this means that as growth continues to materialise on the bottom line, share will continue to rise, but the very rapid upwards moves are now well behind us.
• Both Samsung and SK Hynix are weak today, but I suspect that this has more to do with a resumption of hostilities in the Middle East than a failure of TSMC to beat revenue forecasts.
• Hence, the next time there is a ceasefire, I expect them to bounce right back to where they were.
• However, TSMC is no longer the bargain that it was at 2026 24.4x PER, as it is now more expensive than Nvidia (23.4x 2026 PER), which I would argue is a more focused investment in the AI Boom.

China: Inference at the edge.

• Skyrocketing prices for renting compute have driven many companies to look for ways of reducing their token costs, which is driving them into the arms of the Chinese, who are shaping up to take a significant share of the market.
• For the last couple of years, China has been growing in the open source marketplace and now has almost completely taken it over, with Meta now an irrelevance, having once been the standard.
• This is because Chinese models are now almost always made available with their model weights and sometimes with the training dataset as well.
• Having access to the model weights means that one can download and use a Chinese model on one’s own infrastructure in its final, fine-tuned version.
• Running a model on one’s infrastructure reduces the cost from $x per million tokens to the cost of the hardware and the electricity to run it.
• It also has the benefit of being able to control where one’s data goes and who has access to it, which is a major concern when accessing Chinese models via an API.
• This is the ideal use case for anyone looking to optimise AI costs and has been the main reason that RFM Research has concluded that a large proportion of AI inference will end up running on edge devices or private clouds.
• Many of the best Chinese models will run on a series of Mac Mini’s chained together, meaning that for start-ups, AI has become much more accessible.
• The APIs offered by the Chinese model makers are also much cheaper, meaning that those who are willing to take the data risk can greatly reduce their costs.
• For example, Claude Opus 4.8 costs $25 per million output tokens while GLM5.2 from Z.ai costs around $3.5 per million output tokens according to OpenRouter.
• Part of the reason for this is that Chinese models use far fewer resources than Claude when running inference, but also that the Chinese are deliberately trying to gain market share.
• I suspect that as the market matures, there will be a significant number of tasks where frontier-level performance is not necessary to perform the task at hand adequately.
• Whether this segment becomes the bulk of revenue and profits or a high-volume and low-margin commodity business of little value remains to be seen, but I think that the majors ignore it at their risk.
• This strategy is classic China and looking at what has happened in solar and is happening in automotive, I would be somewhat concerned if I were a frontier lab selling API access at premium prices.
• So far, they have made no move to address this issue but that may change.