Welcome back! Today’s stories follow the money and the power in AI: a founding voice walks away from a giant after losing a fight over the roadmap, new hardware tries to sell AI in a box to anyone who can plug in a rack, investors start worrying less about hype and more about your power bill, and a simple tool quietly turns weekend hustles into live websites. It's a snapshot of AI expanding from research labs to commercial grids and solo businesses, highlighting the contrast between its growth and its practical foundation.
In today’s Generative AI Newsletter:
Yann LeCun exits Meta, calls LLMs dead.
NVIDIA unveils Rubin rack to undercut Blackwell.
Investors warn AI buildout could fuel inflation.
Durable builds small-business websites in minutes.
Latest Developments
When the architect speaks, people listen. Yann LeCun, one of the most influential figures in modern AI, is leaving Meta after more than a decade and is publicly questioning the company’s direction. In a candid interview with the Financial Times, Meta’s outgoing chief AI scientist accused the company of manipulating benchmarks, sidelining research leadership, and doubling down on a technical path he believes will fail. His comments land at a moment when Meta is reshaping its AI org and betting heavily on a new generation of leaders.
What LeCun said:
Benchmarks: Llama 4 results were “fudged a little bit,” using different models to boost scores.
Leadership: Alexandr Wang was called “young” and “inexperienced” in research leadership.
Fallout: Mark Zuckerberg “lost confidence” and sidelined the GenAI team, according to LeCun.
Philosophy: Large language models are a “dead end” for superintelligence, he said.
LeCun said the fallout from Llama 4 led to departures and more exits ahead. “A lot of people have left, a lot of people who haven’t yet left will leave,” he told the FT. On reporting to Wang after Meta’s $14B Scale AI deal, LeCun was blunt. “You don’t tell a researcher what to do. You certainly don’t tell a researcher like me what to do.” He also criticized Meta’s hiring push, saying many new recruits are “completely LLM pilled.” LeCun is now launching Advanced Machine Intelligence Labs, where he will serve as executive chair while Alex LeBrun acts as CEO. The startup will focus on world models trained on video and physical data, not text alone.
Special highlight from our network
Around 80% of AI projects fail. Messy, inconsistent, or hard-to-find data.
This new O’Reilly report shows how to fix that with structured context: clear metrics, solid metadata, and built-in governance. Companies like Walmart and Block are already using this approach to scale safely.
If you want AI that’s accurate, explainable, and trusted, start with this 10-step roadmap.
Special highlight from our network
GenAI Academy presents Vibe Sessions: live, practical, beginner-friendly workshops with Samuel Cummings.
Jan 12: Design Vibes (Image & Video)
Jan 14: Biz Vibes (Business apps)
Jan 16: Web Vibes (Web apps)
Join one or all. 9 to 10 AM Pacific Time.
At CES 2026, NVIDIA changed how companies buy AI as a full rack that can be plugged in and scaled. It unveiled a new platform called Vera Rubin, which is a six-chip stack designed for AI processing. The stack bundles the Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, and Spectrum-6 into what it labels a ready-made AI supercomputer. Selling the whole stack gives NVIDIA more control over pricing, standards, and upgrade timing in the next AI data center cycle.
What's new and what to keep an eye on:
Package: The flagship NVL72 packs 72 GPUs and 36 CPUs built to run like one machine.
Cost: It claims up to 10x lower inference token cost and 4x fewer GPUs to train MoE models than Blackwell.
Control: Jensen Huang introduced a unique data control system based on innovative formats and system techniques.
Scale: NVIDIA adds a KV-cache layer called Inference Context Memory Storage for longer chats and faster replies.
Buyers benefit from improved speed, enhanced security, and reduced integration challenges. Nvidia also sells this as a safety story with rack-level security across compute and infrastructure. However, the downside is that the platform can become restrictive, making changing suppliers akin to rewiring a factory. Rubin is scheduled for release in the second half of 2026 and will be accessible through partners. Competitors will need to compete for success in the market to sell a full rack that operators can trust and afford.
At the start of 2026, the effects of the AI boom are beginning to show up outside stock charts. After a year of record spending by large tech companies, the infrastructure required to run AI systems is expanding rapidly. New data centers, chip factories, and power connections are being built at a pace utilities and suppliers did not plan for. What was once a background investment cycle is now starting to touch energy bills, supply chains, and everyday costs.
Where the pressure is coming from:
Data Center Buildout: Analysts estimate AI data center spending could reach $4T by 2030.
Electricity Demand: AI facilities are driving sharp increases in power consumption.
Chip Costs: Memory and advanced chip prices are rising as inventories tighten.
Policy Risk: Investors warn central banks may delay cuts or restart rate hikes.
The strain is already visible. Oracle disclosed sharply higher spending tied to infrastructure expansion, sending its shares lower. Broadcom warned that component costs are rising faster than expected. HP says memory prices could hurt margins later in 2026. Utilities and regulators are flagging that power demand from AI is growing faster than planned, forcing upgrades that take years to complete. Power plants, transmission lines, chips, and skilled labor cannot scale overnight. As those limits are reached, the costs surface elsewhere.
Durable is an all-in-one tool that helps you get a small business website more quickly. You answer a few questions, it generates a working website that you edit with a drag and drop option. It also includes a blog, contacts and invoicing, so you can publish, capture leads and send bills in one place.
Core functions (and how to use them):
AI site draft: Generate a homepage, services page, and contact page, then swap in your real pricing and photos.
Page flow: Build simple options for Home, Services, About and Contact. Put one clear action on every page.
Copy rewrites: Rewrite your headline, service blurbs, and FAQs to match your actual pricing, turnaround time, and guarantees.
Blog setup: Create 5 posts that answer the buyer and a basic invoice for a project or hourly work and track what’s sent vs paid.
Keep leads organized: Save new inquiries as contacts, tag them (hot lead, follow-up, repeat customer), and stop losing people in your inbox.
Try this yourself:
Pick one service you could sell this week. Generate a Durable site, then change only 3 things: headline to say what you do and who it is for, price anchor “Starts at ₹X or $X” and one contact button. Next, create an invoice template for that same service and send it to yourself as a test.
