Oct 16, 2026

TSMC Posts Record Profit as Meta and Microsoft Unveil Giant New AI Models

TSMC posted a record NT$763 billion quarterly profit on surging AI chip demand, while Meta released an 816B open-weight model and Microsoft finished training MAI-Base-2.

Listen to this update

The AI boom delivered two kinds of proof on Thursday. Taiwan Semiconductor Manufacturing Co. (TSMC), the contract chipmaker behind most of the world's advanced AI processors, reported its biggest quarterly profit ever. Within hours, Meta and Microsoft each revealed some of the largest AI models either company has built, one of them free to download.

TSMC: a sixth straight record quarter

TSMC said net income for the July–September quarter reached NT$763.1 billion, roughly US$23.8 billion, up about 69% from a year earlier and ahead of the roughly NT$746 billion analysts had expected. Revenue rose around 50% to NT$1.49 trillion, or about US$46 billion. It was the company's sixth consecutive record quarterly profit.

Profitability stayed high. Operating margin came in at 58.7%, above the top of the company's own guidance. Gross margin was 66.4%, inside its forecast range, with reports pointing to early costs of ramping up its newest 2-nanometer process as a modest drag.

The mix shows where demand is coming from. High-performance computing, the category that covers AI accelerators and data-center chips, made up 68% of revenue, while smartphones accounted for 21%. Chips made on 7nm-class processes or better produced 79% of wafer revenue, with 3nm alone at about a third. The newest 2nm node is already contributing roughly 8%, according to one breakdown of the results.

Looking ahead, TSMC guided fourth-quarter revenue to between $52.3 billion and $54.3 billion, about a 14% jump from the third quarter. The company said capital spending for 2026 through 2028 will be significantly higher than in the previous three years, and reports put this year's budget at $60 billion to $64 billion. Analysts who follow the company say TSMC has made clear that demand for AI chips still exceeds what its factories can supply.

In the U.S., construction of TSMC's second Arizona fab is complete and equipment is being installed, with mass production now expected in the third quarter of 2027, slightly ahead of schedule. A possible Texas project is being evaluated but has not been approved.

The record quarter did not fully reassure markets. TSMC's U.S.-listed shares and Nvidia each slipped more than 3% during a broader selloff in chip stocks, a sign of how much good news investors had already priced in.

Meta releases Muse Spark Open

Meta put out Muse Spark Open, an open-weight version of its Muse Spark model. It is a sparse mixture-of-experts system with about 816 billion total parameters, of which roughly 42 billion are active for any given token. It takes text, images, audio and video as input, produces text, and supports a context window of up to one million tokens.

Meta pitches the model at long-running agent work, such as coding agents that read a large codebase, plan changes, call tools and fix their own failed steps. The company calls it "the most capable US open-weight model to date" and says its internal benchmarks show it beating China's GLM 5.3 and Kimi K3 on agentic and visual-reasoning tasks. Those results have not been independently verified.

The weights are on Hugging Face under Meta's Muse Model License, and a long list of cloud providers already host the model. Running it yourself takes serious hardware. The full-precision checkpoint is around 1.5 to 1.6 terabytes. Meta's compressed NVFP4 version can run on two Apple M5 Ultra Mac Studios with 512GB of unified memory each, or four 256GB units linked over Thunderbolt 5. Community quantizations shrink it further, down to versions that run on a single 512GB Mac.

Microsoft finishes training MAI-Base-2

Microsoft AI said it has completed pre-training MAI-Base-2, a mixture-of-experts foundation model with about 2.8 trillion total parameters and around 80 billion active. Microsoft says the run used 16,000 Nvidia GB200 GPUs over 38 days in August and September, on a training mix of about 50 trillion tokens.

Microsoft's headline claim is efficiency. It says the model is 4 to 12 times more efficient in pre-training compute than DeepSeek-V4-Pro-Base, and that it scored better than every open-weight model it tested on held-out evaluations. Microsoft AI chief Mustafa Suleyman went further, telling Business Insider the model beat a comparable OpenAI model using about a quarter of the compute. All of these comparisons rely on Microsoft's private tests.

You can't use MAI-Base-2 yet. It is now in post-training, and Microsoft plans to license the weights to trusted partners through its Foundry platform. Work on MAI-Base-3 has already begun.

Why it matters

Taken together, Thursday's news shows both ends of the AI supply chain moving fast. TSMC's results suggest spending on AI hardware is still rising, and the models announced by Meta and Microsoft are the kind of workloads that spending pays for. Meta's release also adds a powerful U.S. option to an open-weight field that Chinese labs have led for much of the past year, while Microsoft's model signals that it wants to rely less on outside labs for its own AI.