AI Week, Jan 26–Feb 1, 2026: Custom chips, Amazon talks on up to $50 billion for OpenAI and a million satellites
From January 26 to February 1, Microsoft launched its Maia 200 inference chip, Brussels opened DMA proceedings on Google, Amazon cut about 16,000 jobs while discussing a stake of up to $50 billion in OpenAI, Apple bought Q.ai, and SpaceX asked the FCC to approve up to 1 million orbital data-center satellites.
Written on November 15, 2026, covering news from Jan 26–Feb 1, 2026.
Microsoft showed a new in-house chip, Amazon cut staff while discussing a large OpenAI investment, Apple bought an audio AI startup, and SpaceX filed for permission to put data centers in orbit. Regulators in Brussels and Beijing also made moves on access to platforms and chips.
Monday, January 26
Microsoft introduced Maia 200, an AI accelerator designed for inference. It is the successor to Maia 100, which Microsoft announced 2 years earlier and never offered to cloud customers for rent.
Microsoft lists these main specifications:
- Built on TSMC’s 3-nanometer process, with more than 140 billion transistors per chip.
- More than 10 petaFLOPS at FP4 and more than 5 petaFLOPS at FP8, within a 750W SoC TDP.
- 216GB of HBM3e at 7 TB/s, plus 272MB of on-chip SRAM.
- 2.8 TB/s of bidirectional scale-up bandwidth per accelerator, with collective operations across clusters of up to 6,144 accelerators.
Microsoft claims Maia 200 delivers 3 times the FP4 performance of Amazon’s third-generation Trainium and higher FP8 performance than Google’s seventh-generation TPU. It also says the chip offers 30% better performance per dollar than the newest hardware already in its fleet. Scott Guthrie, Microsoft’s executive vice president for cloud and AI, called it “the most efficient inference system Microsoft has ever deployed”.
The networking uses standard Ethernet rather than the InfiniBand standard. CNBC noted that Nvidia sells InfiniBand switches following its 2020 acquisition of Mellanox. Each tray connects 4 Maia accelerators directly to one another, without switches.
The chip is running in Microsoft’s US Central region near Des Moines, Iowa. US West 3 near Phoenix, Arizona, is next, and more regions will follow. Microsoft says Maia 200 will serve several models, including OpenAI’s GPT-5.2, and will support Microsoft Foundry and Microsoft 365 Copilot. The superintelligence team led by Mustafa Suleyman will use it. Microsoft says that team will apply it to synthetic data generation and reinforcement learning.
On availability, Guthrie said there will be “wider customer availability in the future”. In the meantime, Microsoft is previewing a Maia SDK. The preview includes PyTorch support, a Triton compiler, low-level programming in NPL, and a simulator with a cost calculator. Microsoft also says models were running on Maia 200 within days of the first packaged parts arriving.
Tuesday, January 27
The European Commission opened 2 sets of specification proceedings under the Digital Markets Act. Their purpose is to spell out how Google should meet 2 of its obligations. EUobserver explained that the proceedings do not, in themselves, investigate possible infringements.
The first set covers Article 6(7), which requires free and effective interoperability with features controlled by Android. This round focuses on features used by Google’s own AI services, such as Gemini. The Commission plans to specify how rival AI providers can get equally effective access to the same features.
The second set covers Article 6(11), which requires Google to share anonymised ranking, query, click and view data from Google Search with rival search engines on FRAND terms. The Commission will look at the scope of the data, the anonymisation method, the conditions of access, and whether AI chatbot providers qualify.
Teresa Ribera, the Commission’s vice-president for competition, said: “We want to maximise the potential and the benefits of this profound technological shift by making sure the playing field is open and fair, not tilted in favour of the largest few”. Clare Kelly, senior competition counsel at Google, told EUobserver: “Android is open by design, and we’re already licensing search data to competitors under the DMA”. She added that Google is concerned further rules “will compromise user privacy, security, and innovation”.
The Commission will send preliminary findings within 3 months and close the proceedings within 6 months. Non-confidential summaries of the preliminary findings and proposed measures will be published so that third parties can comment. The Commission also said the proceedings do not limit its power to find non-compliance, which can include fines.
Wednesday, January 28
Amazon is cutting about 16,000 jobs. Senior vice president Beth Galetti announced the cuts in a blog post that reproduced her message to employees. CNBC described them as corporate roles and as Amazon’s second round of mass job cuts since October.
Galetti addressed the question of whether broad reductions will now come every few months: “That’s not our plan”. She said some teams had finished their organizational changes in October and others had only just completed them. She also said Amazon will keep hiring in strategic areas.
Staff based in the US will have 90 days to find another role inside the company. After that, they will be offered severance, outplacement services and health insurance benefits.
Thursday, January 29
Apple confirmed it has acquired Q.ai, an Israeli startup working on AI for audio. Apple did not disclose the price. Reuters, citing a source familiar with the matter, reported that the deal valued Q.ai at about $1.6 billion.
According to Apple, Q.ai worked on machine learning that helps devices understand whispered speech and improves audio in difficult environments. Last year the startup filed a patent application to use “facial skin micromovements” to detect words that are mouthed or spoken. About 100 employees will join Apple, including CEO Aviad Maizels and co-founders Yonatan Wexler and Avi Barliya.
Maizels previously sold PrimeSense to Apple in 2013. PrimeSense’s technology was integrated into Apple’s Face ID sensor starting in 2017. Johny Srouji, Apple’s senior vice president of hardware technologies, said: “We’re thrilled to acquire the company, with Aviad at the helm, and are even more excited for what’s to come”. Q.ai’s backers included Matter Venture Partners, Kleiner Perkins, Spark Capital, Exor and GV.
CNBC noted that some investors have urged Apple to make a large AI acquisition, and that Apple more commonly buys smaller companies with specific technologies. In July, Tim Cook said: “We’re very open to M&A that accelerates our roadmap”.
The same day, CNBC confirmed talks under which Amazon could put as much as $50 billion into OpenAI. The Wall Street Journal first reported that figure. A source told CNBC that Sam Altman and Andy Jassy have been talking directly. The same source said the terms are still fluid, the final amount could change, and a term sheet could be signed in the coming weeks.
According to that source, the full OpenAI round could reach about $100 billion. It could close in 2 parts: strategic investors such as Amazon, Microsoft and Nvidia first, with SoftBank and others following. SoftBank is in talks to invest up to $30 billion. CNBC reported earlier that a final deal could include an agreement for OpenAI to use Amazon’s AI chips. OpenAI was valued at $500 billion in October, after a secondary share sale.
Amazon has invested billions in Anthropic since 2023. AWS remains Anthropic’s primary cloud and training partner. Amazon expects capital expenditures of $125 billion for 2026.
Friday, January 30
SpaceX filed a request with the FCC to launch up to 1 million solar-powered satellites that would serve as AI data centers. The filing calls the plan “a first step towards becoming a Kardashev II-level civilization”.
According to the filing, the satellites would:
- Operate at altitudes between 500 km and 2,000 km.
- Communicate with each other through laser links.
- Rely on radiative cooling in space rather than water-based cooling.
SpaceX projects that launching 1 million tonnes of satellites per year, at 100 kW of compute per tonne, would add 100 gigawatts of AI compute capacity annually. The filing gives no deployment timeline. Musk replied on X: “I thought we’d start small and work our way up”.
The Verge argued that the 1 million figure is unlikely to be approved outright and is probably an opening position for negotiation. The FCC recently approved 7,500 more Starlink satellites for SpaceX but deferred a decision on the remaining 14,988. Analyst Tim Farrar called orbital data centers a “Rorschach test” for investors. Kimberly Siversen Burke of Quilty Space said that, as a near-term source of revenue, they “remain speculative”. In a story updated on February 2, Via Satellite reported that SpaceX had acquired xAI.
Also this week
On January 30, TrendForce reported, citing Reuters, that China had given DeepSeek conditional approval to buy Nvidia’s H200 chips, with the regulatory terms still being finalized. Reuters had earlier reported that ByteDance, Alibaba and Tencent were authorized to buy more than 400,000 H200 chips between them. A source said China’s National Development and Reform Commission is setting the terms.
Bloomberg reported on January 29, citing Jensen Huang in Taipei, that Nvidia had not yet received any H200 orders from China. According to Reuters, any DeepSeek purchase could also face scrutiny from US lawmakers. The Information reported that DeepSeek plans to unveil its V4 model in mid-February.
What to Watch
- OpenAI term sheet: whether Amazon’s term sheet is signed in the coming weeks, and whether the final terms include using Amazon’s AI chips. My read is that this would be the clearest sign of whether Amazon is spreading its AI bets beyond Anthropic.
- EU preliminary findings: the draft measures for Google due within 3 months, especially whether AI chatbot providers qualify for access to Search data.
- Maia 200 rollout: the US West 3 deployment, and whether the “wider customer availability” Guthrie promised arrives. I think that matters more than the benchmark claims, since Maia 100 never reached rental customers.
- DeepSeek: the final terms on H200 purchases and the reported mid-February launch of V4.
This article was written by AI from the sources listed on this page. Figures, names and quotes were checked against those sources by software before publishing.