Trying out a new schedule this week! If you have strong opinions on Friday over Saturday, leave a note in the comments!
Flash Forward
Fast (pun intended) on the heels of its flagship Pro model, Google launched Gemini 3 Flash, an impressively high-scoring model given its reduced cost and latency.
The big picture:
Gemini 3 Flash outperforms its predecessor by significant margins (33.7% vs 11% on Humanity’s Last Exam) while being positioned as a “workhorse model” at $0.50 per million input tokens.
The Flash model series has become Google’s most popular offering, suggesting developers prioritize deployment economics over cutting-edge performance (though there is a difference between dollars spent and tokens spent).
Following reports of OpenAI’s internal “Code Red” memo as Google gains consumer market share, both companies are locked in a rapid-fire release cycle that’s processing over 1 trillion tokens per day through Google’s API alone.
Elsewhere in frontier models:
Mistral launches Mistral OCR 3, featuring improvements in processing forms, scanned documents, complex tables, and handwriting, priced at $2 per 1,000 pages.
Nvidia launches Nemotron 3, a family of AI models using a hybrid mixture-of-experts architecture and the Mamba-Transformer design, in 30B, 100B, and ~500B sizes.
The Allen Institute for AI launches Bolmo 7B and Bolmo 1B, claiming they are “the first fully open byte-level language models“, built using its Olmo 3 models.
And OpenAI releases GPT‑5.2-Codex, with improvements on long-horizon work through context compaction, stronger performance on large code changes, and more.
Elsewhere in the FAANG free-for-all:
Google is working on a new initiative to make its AI chips run PyTorch better and is collaborating closely with Meta, as the two discuss Meta using more TPUs.
Meta is developing a new image and video-focused AI model codenamed Mango, expected to be released in H1 2026 along with its new LLM dubbed Avocado.
Amazon’s AGI team lead, Rohit Prasad, is leaving at the end of 2025; AWS SVP Peter DeSantis will lead a group combining AI, silicon, and quantum computing teams.
And Google brings its vibe coding tool Opal to Gemini on the web, letting users create their own custom AI-powered mini apps, called Gems, which can be reused later.
Cracks in the Foundation
A wave of construction delays, funding pullbacks, and market skepticism is hammering AI infrastructure companies, erasing tens of billions in market value and raising urgent questions about whether the frantic race to build data centers can deliver returns that justify the massive capital investments.
Why it matters:
CoreWeave has lost $33 billion in value over six weeks - a 46% plunge - while data center landlord Fermi’s stock nearly halved after Amazon withdrew $150 million in construction funding, and Oracle pushed back some OpenAI data center completions to 2028.
The financial stress appears to be mounting across the sector. CoreWeave operates on razor-thin 4% margins while paying double that in interest on its debt - and even profitable giants like Oracle and Broadcom saw double-digit stock declines after disappointing investors on spending timelines.
As one investor put it, “every single AI company on the planet is saying if you give me more compute I can make more revenue” - but the market is no longer taking that promise on faith.
Elsewhere AI infrastructure:
A study finds that AI’s 2025 power demand could hit 23GW, exceeding 2024 Bitcoin mining levels, with carbon emissions reaching 32.6M to 79.7M tons compared to NYC’s 50M.
Hut 8 partners with Fluidstack to build an AI data center in Louisiana for Anthropic, backed by a 15-year, ~$7B lease starting with 245MW of computing capacity.
Nvidia acquires SchedMD, the developer of Slurm, an open-source AI workload management system, and will continue distributing it on an open-source basis.
And Oracle signed ~$150B of data center leases in the three months ending November 30, raising its total data center and cloud capacity commitments to $248B.
Elsewhere in AI anxiety:
The US FTC is investigating Instacart’s AI pricing tool after a study showed different prices for the same items at the same stores, causing CART stock to drop over 8%.
Hollywood is divided over AI adoption, with some opposing all AI use while directors like Timur Bekmambetov are exploring ways to embrace AI tools.
Copywriters report that generative AI has led to their work being used for training, layoffs, plummeting wages, and freelancers losing clients.
And Elizabeth Warren and two other Democratic senators wrote to Google, Microsoft, Amazon, Meta, and others about AI data centers’ role in rising electricity bills.
China’s Secret Chip Lab
China has secretly built a working prototype of an advanced chip-making machine that could eventually break Western dominance in semiconductor manufacturing, years ahead of Western estimates.
Between the lines:
The prototype, completed in early 2025 in a high-security Shenzhen lab, is operational and generating EUV light but hasn’t produced working chips yet. China aims for 2028, though experts say 2030 is more realistic.
This is China’s “Manhattan Project” for semiconductors, coordinated by Huawei and involving thousands of engineers working under strict secrecy. Some recruits were given fake IDs to conceal their identities, and the effort involves salvaging parts from older machines and exploiting secondary markets to bypass export controls.
Yet despite this breakthrough and massive AI investment battles between the U.S. and China, recent analysis reveals China’s AI chip capabilities are falling further behind, not catching up - Nvidia’s chips are currently five times more powerful than Huawei’s best offerings, expanding to seventeen times by 2027.
On the other hand, America’s concentrated $350+ billion bet on AI contrasts with China’s diversified strategy spreading investments across electric vehicles, batteries, solar panels, and other manufacturing sectors - raising the prospect that the US could win the AI battle but lose the broader economic war.
Elsewhere in Chinese AI:
At least six Chinese AI companies, including chip firms Montage and GigaDevice, are set to IPO in Hong Kong in the coming weeks, aiming to raise up to $1B.
Chinese AI chipmaker MetaX’s shares jumped as much as 755% in its Shanghai debut, after a heavily oversubscribed ~$585.8M IPO, pushing its market cap past ~$42B.
Nvidia told its Chinese clients that it is evaluating adding production capacity for its H200 chips after orders exceeded its current output level.
And Taiwan opened its largest AI supercomputing data center, with Nvidia’s Blackwell chips, in a major effort for its push for sovereign AI and chip industry innovation.
Elsewhere in AI geopolitics:
The UK AI Security Institute reports that AI models are rapidly improving at potentially dangerous biological and chemical tasks and show fast jumps in self-replication capabilities.
a16z released a nine-pillar AI policy framework for Congress, including restricting AI use for under-13s without parental consent and keeping a role for states.
A look at how David Sacks and AI adviser Sriram Krishnan overcame MAGA opposition to Trump’s AI executive order by talking to lawmakers and modifying its language.
And Trump’s AI executive order directs federal agencies to look into withholding federal funding from states passing “the most onerous and excessive laws” around AI.
Skills Without Borders
A lightweight “skills” format introduced by Anthropic in October for teaching AI assistants repeatable workflows is rapidly becoming an industry standard, with OpenAI implementing it within ChatGPT and Codex - signaling a rare moment of interoperability in the competitive AI landscape.
The big picture:
Skills represent a deceptively simple but powerful specification: just a folder containing a Markdown file with instructions and optional resources that any AI system with filesystem access can implement. They’re more accessible than complex agent frameworks, more portable than platform-specific features, and more standardized than custom prompts.
Anthropic is betting on an ecosystem play, launching a partner directory with pre-built skills from Notion, Canva, Figma, and Atlassian while adding enterprise management features that let admins provision skills organization-wide.
The speed of OpenAI’s adoption (appearing just weeks after Anthropic’s October launch) and the emergence of community repositories suggest skills are solving a real pain point. The question now is whether this informal standard will be formalized - perhaps through initiatives like the new Agentic AI Foundation - or remain an organic speci that evolves through practice rather than governance.
Elsewhere in OpenAI:
OpenAI has held preliminary discussions with investors about raising tens of billions of dollars, potentially as much as $100B, at a valuation of around $750B.
OpenAI hires former UK chancellor George Osborne to lead OpenAI for Countries, the global expansion arm of its $500B Stargate data center initiative.
OpenAI opens app submissions for review and publication in ChatGPT, where users can discover or search for apps in a new app directory.
OpenAI debuts FrontierScience, a benchmark to measure models’ expert-level scientific capabilities with 700+ questions, finding GPT-5.2 is its strongest model.
And OpenAI updates ChatGPT Images to generate images up to 4x faster and edit uploads more precisely, adding a new section in ChatGPT’s app and website for images.
Things happen
Tencent names former OpenAI researcher Yao Shunyu as chief AI scientist. xAI’s lack of enterprise sales experience is holding back potential customers. AI image generators now mimic phone camera traits to avoid the uncanny valley. Q&A with Sam Altman on OpenAI’s “code red” call, enterprise strategy, and IPO plans. Claude ran a vending machine and lost $1,000+ after dropping prices to zero. Mira Murati’s lab makes Tinker generally available. YouTube terminates two channels that used AI to create fake movie trailers. Google launches CC, an AI assistant that delivers personalized daily briefings. Mark Zuckerberg’s all-in AI bet includes suffocating micromanagement, sources say. Google removes AI videos of Disney characters after cease-and-desist letter. Amazon rolls out Ask This Book to Kindle; authors can’t opt out. Twenty-four companies join the US Genesis Mission to boost AI for scientific discovery. The AI boom is delaying municipal projects as workers shift to data centers. Anthropic exec forces AI chatbot on gay Discord community, members flee. Chinese AI infrastructure companies surge despite Trump’s tariffs. Trump’s AI Executive Order could invalidate GOP state laws targeting “Big Tech censorship.” GOP governors push ahead with their own AI bills despite Trump’s EO.
Publish on Fridays please rather than Saturdays! You quite often have links I miss in my trawling for my Friday-published newsletter https://thisweekinnlp.substack.com/.
This piece really made me think about the practical side of AI development. I especially liked your point about developers prioritising deployment economics over cutting-edge performance with models like Gemini 3 Flash. It’s such an insightful observation, becouse it highlights the real-world considerations beyond just benchmark scores. Really great roundup!