AMD Posts a Record Quarter on Data Center Chips as OpenAI Buys Pierre Computer and Tencent Ships Hy4 With Computer Use

AMD Books a Record Quarter as Data Center Sales Nearly Double

AMD reported third-quarter 2026 results after the market closed on Tuesday, and the numbers put the company squarely in the AI infrastructure conversation. AMD’s third-quarter 2026 earnings release shows record revenue of $13.38 billion, up 45% from $9.25 billion a year earlier and up 16% from the second quarter. GAAP net income was $2.43 billion, or $1.46 per diluted share, and non-GAAP earnings per share were $2.09, up 74%. The Data Center segment, which includes EPYC server processors and Instinct GPUs, delivered a record $8.5 billion, up 97% year over year, which AMD tied to strong EPYC demand and the launch of its Instinct MI450 Series GPUs.

The rest of the business was mixed. Client and Gaming revenue fell 8% to $3.7 billion, with gaming down 40% to $773 million on lower semi-custom and graphics sales, while Embedded grew 29% to $1.1 billion. The bigger signal was the outlook: AMD expects fourth-quarter revenue of about $17 billion, plus or minus 3%, which it says would be roughly 66% growth from a year earlier and a 27% jump from the third quarter. CNBC’s report on the AMD results noted that revenue and adjusted earnings beat LSEG consensus estimates of $13.09 billion and $1.95 a share, and that the stock jumped 7% in extended trading. Chair and CEO Lisa Su framed the quarter this way:

“We delivered a record quarter with revenue exceeding the high end of our guidance as Data Center segment sales nearly doubled year-over-year,”

Su also said AMD expects growth to accelerate further as it scales production of its Helios rack-scale systems and its “Venice” server processors.

OpenAI Moves to Buy Pierre Computer and Its Code Storage for Codex

OpenAI will acquire Pierre Computer, a developer tools startup whose technology already runs inside Codex, according to Axios’s exclusive on the OpenAI and Pierre Computer deal, which says OpenAI told the outlet directly. OpenAI plans to fold Pierre’s Code Storage product into Codex. Financial terms were not disclosed, and Pierre’s open-source projects will remain open source. As of Wednesday morning, we had not found a separate announcement on OpenAI’s or Pierre’s own sites, so the details here rest on the Axios report; the deal’s price and closing date are TBD.

Pierre’s site describes Code Storage as a Git-based distributed storage system with APIs designed for machines, aimed at products where agents create, branch and push code at high volume. Axios places the deal in a wider fight over the place where software lives, noting that GitHub has faced reliability problems as AI coding has taken off and that several rivals have launched competing products this year. The deal comes days after OpenAI described Pando, its in-house Git backend, which we covered on Tuesday. Axios summed up the strategic direction plainly:

OpenAI is moving beyond helping developers write code toward controlling more of the infrastructure where that code is stored, reviewed and shipped.

OpenAI Shows Codex and ChatGPT Doing Physical Engineering Work

OpenAI also published OpenAI’s post on accelerating engineering for the physical world, which describes how its models help with mechanical design, electrical engineering and simulation. In one test, two team members used GPT-6 Astra to design a small device for talking to Codex, covering the enclosure, the printed circuit board, the firmware and the user interface, without direct human edits to the design. OpenAI says the work took under a week as a side project and that a contract manufacturer built a small batch of prototypes in under a month. Its hardware teams now use Codex and ChatGPT for CAD models, manufacturing drawings, simulation setup and design review, and OpenAI says hardware-team token usage has grown 23-fold since February.

The post also announces a research collaboration with Mazda on aerodynamics and crash-safety simulation, an expanded collaboration with Chip Ganassi Racing on turning 3D scans into editable CAD, and an internal BenchCAD 2.0 result in which a newer internal model averaged 0.875 against 0.594 for GPT-6 Astra. OpenAI was candid about where the device test fell short:

Astra did not consistently account for the manufacturer’s production limits, tolerances, or details required for assembly.

Tencent Releases Hy4 With Vision and Computer Use

Tencent on Wednesday announced Hy4, the newest model in its Hy family, and promised to publish the model weights and system cards “in the coming days.” Tencent’s Hy4 launch announcement says the model adds visual understanding and computer use for the first time, so it can work across command-line and graphical interfaces, with specialized training on software such as Blender and CAD. Hy4 is available now on Tencent Cloud TokenHub and in Tencent products including WorkBuddy, CodeBuddy and Yuanbao, with API access through third-party platforms such as OpenRouter. Pricing is unchanged from Hy4 Preview at about $0.90 per million input tokens and $2.70 per million output tokens, and TokenHub is offering 50% off through December 31, 2026.

Tencent reports that, in its internal evaluations, Hy4 cut its hallucination rate from 12.95% in the preview to 7.2%, and that it scored highest in a blind evaluation of 579 financial tasks judged by 213 financial professionals. Those are company-run tests. On safety, Tencent says Hy4 is trained on defensive security data and that its Xuanwu and Keen labs found no significant offensive cyber concerns in the areas they evaluated. The release states:

However, offensive materials related to exploitation are explicitly excluded from the training dataset.

Oracle Reportedly Cancels a $7 Billion Chip Lease With Tencent

Oracle has cancelled a multibillion-dollar contract to lease AI chips to Tencent after pressure from the Trump administration, the Financial Times reported on Wednesday, according to The Business Times’ Reuters account of the Financial Times report. The FT said US officials approached Oracle to raise concerns about the deal, citing people familiar with the matter, and that officials in Guangdong, Tencent’s home province, asked Tencent to support domestic Chinese chips. The FT reported in October that the five-year lease, worth an estimated $7 billion, would have given Tencent access to about 100,000 advanced AI chips across Oracle data centers in Southeast Asia.

This is a report, not a confirmed announcement. Reuters said it could not immediately verify the FT story, and Oracle, Tencent and the White House did not respond to its requests for comment. The story lands the same day Tencent launched Hy4, a reminder that Chinese labs continue to ship competitive models while access to the most advanced foreign compute narrows.

Menlo Ventures Puts Enterprise AI Spending at $176.8 Billion

Menlo Ventures released its fourth annual enterprise AI report on Tuesday. Menlo Ventures’ 2026 State of Enterprise AI report estimates enterprise spending on deploying AI at $176.8 billion as of September 2026, up 5.4 times from $33 billion in 2025, against $102.5 billion spent on building models. Anthropic, OpenAI and Google together take 69% of deployment spending, up from 50%. Within enterprise LLM API spending, Menlo puts Anthropic at 55%, up from 40%, OpenAI at 31% and Google at 9%, down from 21%. Coding, at $34.0 billion, and assistants and chatbots, at $25.0 billion, make up 76% of a $77.3 billion application market. On OpenRouter, agents now generate 80% of all tokens.

The estimates come from a bottom-up model plus a Morning Consult survey of nearly 400 US enterprise decision-makers fielded September 8 to 26, 2026, and Menlo is an Anthropic investor, which readers should weigh. Menlo’s GlobeNewswire release on the report added a striking line on open models:

Open-source models now handle approximately one in five tokens but account for only about one in 25 dollars of model spending.

Mistral Shows 133 Agents Finding and Patching a Django Flaw

Mistral AI posted a demonstration on Wednesday in which Mistral Large 4, the open-weight model it calls Le Chonk, was set loose on a recent Django vulnerability. In Mistral AI’s official X post on the Django CVE demo, the company says a swarm of agents mapped the code, tested attack scenarios and confirmed the bug before the model wrote a patch, shown in a video sped up 90 times. Mistral did not name the specific CVE in the post, so that detail is TBD, and this is a vendor demo rather than an independent evaluation. The post reads:

133 agents map the code, hunt for flaws, test attack scenarios and confirm the bug. Then Le Chonk patches it.

A day earlier, Mistral said Mistral Large 4, with 1.05 trillion open weights and a 1 million token context window, had joined the lineup on Open Math OS, a shared research workspace for mathematicians and AI agents, for approved teams.

What We Left Out

An OpenAI staff post teasing “ChatBLT” was read by some on X as a hint at a Thursday model launch. OpenAI has announced nothing, so we treat it as a rumor. Reports that DeepSeek will pre-train its V5 model mainly on Huawei Ascend chips remain unconfirmed by DeepSeek. We also held back a reported funding round for Fal, a reported US delay to SK Hynix tool licenses for its Dalian plant and a reported MIT result on weight-space editing until we can confirm them against primary sources.