The Brief: Frontier Chips & Training Data
Multiple AI labs including Meta and DeepSeek are reportedly building their own chips. Plus: Cursor trained a frontier model on users' data
THE DOWNLOAD
DeepSeek Reported to Be Developing an Inference-Focused AI Chip
Reuters reports, citing three unnamed sources, that DeepSeek has spent about a year designing a chip built for running models rather than training them, and is talking to chip design, foundry, and memory partners. DeepSeek has not commented, and nothing is confirmed.
Why it matters: The decision to build only for inference says a lot. DeepSeek’s own research measured Huawei’s Ascend 910C at about 60% of an Nvidia H100 on inference, and its R2 model reportedly had to fall back to Nvidia hardware for training. With roughly 70% of AI compute demand expected to be inference, a lab that designs both its own models and its own chips could squeeze Nvidia’s China business even if the chip never leaves the country.
Meta to Begin Production of Its Iris AI Chip in September
Meta’s newest in-house AI chip, code-named Iris, goes into production in September, designed with Broadcom and manufactured by TSMC, according to an internal memo obtained by Reuters. Meta plans 7 gigawatts of computing capacity in 2026, doubling to about 14 gigawatts in 2027, on spending of $125B to $145B this year.
Why it matters: Meta is not replacing Nvidia and AMD, it is adding its own chips alongside them. The steadier winner is Broadcom, which designs custom chips for Google, OpenAI, Meta, and ByteDance and holds a $73B backlog of AI orders. Custom chips are on track to reach about 28% of AI server shipments in 2026, growing nearly three times faster than off-the-shelf GPUs.
SK Hynix Raises $26.5B in Record US Share Sale; Shares Open 14% Above Offer
SK Hynix raised $26.5B in the largest first-time US share sale ever by a foreign company, topping Alibaba’s 2014 record, with demand running about seven times the shares on offer. The stock opened roughly 14% above the $149 offer price in Friday’s Nasdaq debut; Seoul shares are up 634% in a year. CEO Kwak Noh-jung told Reuters that 2027 will be the worst supply shortage year in the memory industry’s history.
Why it matters: SK Hynix supplies 56.4% of the high-bandwidth memory that AI chips depend on, and the deal cleared seven times demand in the same week memory stocks sold off sharply. The caution flag: analyst Patrick Moorhead notes that memory makers have sold below cost in past downturns, and the five-year customer contracts meant to prevent a repeat are untested. Chairman Chey Tae-won also said the company is open to selling more US shares.
OpenAI Launches ChatGPT Work and the GPT-5.6 Model Family
ChatGPT Work handles multi-step tasks across Slack, Teams, Google Drive, and SharePoint, builds documents, slides, and simple web apps, and can run tasks on a schedule.
Why it matters: Companies have been hit with surprise bills as AI usage grows, and Terra at half the price of GPT-5.5 answers that complaint directly. Early reviews are mixed: M.G. Siegler argues OpenAI copied the shape of Anthropic’s Claude Cowork but executed it less gracefully, burying the chat experience inside a work app.
xAI and Cursor Release Co-Developed Grok 4.5 at $2/$6 Per Million Tokens
Grok 4.5 launched July 8. It was co-developed with Cursor and trained on what Cursor describes as trillions of tokens of its data, and is available in Cursor from day one. xAI claims it beats Claude Opus 4.8 on coding benchmarks; independent testing ranks it fourth on general intelligence but first at using tools.
Why it matters: This is the cheapest model near the top of the market, albeit not the best one. The post training data itself is the new thing: Cursor trained a model on its users’ coding activity with a single lab, and other coding tools, and the customers whose work fed the training, now have to decide how they feel about that.
Cognition Releases SWE-1.7, a Coding Model Built on Kimi K2.7
SWE-1.7 starts from Moonshot AI’s Kimi K2.7 Code model, runs through Cerebras at 1,000 tokens per second, and is free at the standard tier inside Devin. In a companion post, Cognition reports the original Chinese model completed 87% of tasks that US models refuse, including writing surveillance software, and says its additional training brought that behavior in line with Claude and GPT.
Why it matters: Kimi K2.7 had already been through heavy reinforcement training at Moonshot, yet Cognition’s extra training on its own usage data still produced large gains. Companies with the right data can apparently keep turning open Chinese models into competitive products for a fraction of what training from scratch costs.
.
ByteDance Releases Seedream 5.0 Pro Image Model
The July 8 release adds precise editing tools, layer separation with transparent export, infographic generation, and text rendering in more than 10 languages, at roughly $0.03 per image versus about $0.28 for GPT Image 2.
Why it matters: Seedream is roughly 3x to 9x cheaper depending on resolution, but GPT Image 2 still renders text more reliably, 98.5% versus 89.5% accuracy, so design teams may not save as much as the price list suggests. Hands-on comparisons put Seedream ahead on price and editing features, not image quality.
Thinking Machines Lab Publishes Vision for AI That Extends Human Judgment
The lab published “The Future Worth Building Is Human” on July 10, arguing AI should extend human will and judgment rather than replace it, and that users, not a single lab, should decide how models behave by training them on their own data. The essay rejects the industry’s favorite progress measure, how long an AI can work without supervision.
Why it matters: The essay doubles as the case for Tinker, the company’s product for customers who train and own their own models, and stakes out a position against the fully autonomous agents the other big labs are chasing.
EVENTS
Join us for cold beer and Hot Chips in the Fall!
