ByteDance Wants Its Own Inference Chips

ByteDance chips are the clearest signal today, even if the clearest fact is still that the company has not confirmed any of it. Elsewhere, production AI keeps moving down-stack, into code review, workflow state, and latency plumbing, while one gaming labor story remains notably underconfirmed.

ByteDance chips push toward inference control

ByteDance chips are still a reporting story, not a company announcement. Reuters reported on May 28 that ByteDance is developing custom CPUs for AI infrastructure and inference-heavy workloads, while The Information reported on May 29 that the company is also working on a Groq-like inference chip project and a processor code-named Ada-S. ByteDance’s public site does not confirm either effort.

The reporting is also not perfectly aligned. Reuters framed the work as custom CPUs and said the project was early stage, with ByteDance not responding to a request for comment. Earlier Reuters-linked reporting in February said ByteDance was pursuing an inference-focused AI chip and talking with Samsung about manufacturing, but also said ByteDance denied that description of an in-house chip project. A separate May 26 report tied to Bloomberg said Qualcomm had reached a chip deal with ByteDance and that the company could buy millions of ASICs. Put together, the signal is less “launch” than “large buyer trying to own more of inference economics.”

Source: theinformation.com

Cloudflare puts multi-agent code review in CI

Cloudflare says every merge request on its standard CI pipeline now gets AI code review. In its April 20 post, the company said the system uses a multi-agent review coordinator that classifies each merge request as trivial, lite, or full, then routes work to specialized agents for code quality, security, codex compliance, documentation, performance, and release impact, according to the Cloudflare blog.

The rollout numbers make this more than a demo. Cloudflare said the system handled 131,246 review runs across 48,095 merge requests in 5,169 repositories in its first 30 days. Median review time was 3 minutes 39 seconds, average cost was $1.19, median cost was $0.98, and P99 cost was $4.45. In a separate internal engineering post, Cloudflare said it has 100% AI code reviewer coverage across repos on its standard CI pipeline. The obvious gain is throughput. The less obvious tradeoff is that code review has also been one of the few routine ways teams spread architectural context around.

Source: blog.cloudflare.com

SQLite becomes agent workflow state by default

SQLite has not rebranded itself as an agent framework. Its own docs still describe it as an in-process, serverless, zero-configuration database engine that reads and writes directly to disk files, per SQLite.org. That is exactly why it keeps showing up underneath agent systems: state, queues, checkpoints, and retries are easier to ship when the database is a library, not another service.

The stronger evidence is in the surrounding products. SQLite.ai now markets “SQLite-Agent” as autonomous agents that run from SQLite, with agent memory and MCP tools built in. Other agent products are making the same architectural bet. r8r lists SQLite as embedded storage for an agent-native workflow engine, and Holons says all data lives in local SQLite while its engine creates run rows and pushes nodes onto a task queue. That does not make SQLite an official workflow engine. It does suggest a lot of agent orchestration software is discovering that one file and ACID semantics solve more of the problem than expected.

Source: www3.sqlite.org

Inference stack claims exceed what sources confirm

The specific claim here, 3,000 tokens per second per request, is not supported by the primary sources provided. The strongest canonical source is OpenAI’s engineering post on WebSockets in the Responses API, where the company said it made agent loops 40% faster end-to-end and raised observed inference speed from 65 to nearly 1,000 tokens per second, according to OpenAI.

That still matters, because it points to software-path latency as a real bottleneck: OpenAI attributed the gains to caching, fewer network hops, faster safety checks, and persistent connections. But it is not the same as a verified 3,000 tok/s result on standard GPUs, and there is no safe source here for that number. Useful direction, weak receipts.

Source: openai.com

GTA 6 union claims remain unconfirmed

The underlying labor tension around Rockstar is real, but the specific claim that GTA 6 developers unionized before launch is not confirmed by the sources here. Take-Two’s own materials are about release timing, not unionization: one investor relations statement tied GTA VI to May 26, 2026, and a later results release moved the date to November 19, 2026, per Take-Two.

The closest sourced labor reporting is Bloomberg from November 5, 2025, which said Rockstar disputed allegations that layoffs were meant to disrupt a unionization attempt and said the fired employees were leaking company secrets. That is evidence of conflict, not evidence that a union drive succeeded. With no Rockstar, Take-Two, union, or regulator source confirming a unionization event, this stays in the category of notable allegation, not established fact.

Source: bloomberg.com

A fair amount of today’s AI news comes with an asterisk. The infrastructure shift is real; the sourcing still matters.