OpenAI and Anthropic kept things relatively quiet. Their competitors and financial backers? Not so much.

Reports showed NVIDIA is investing $26 billion in open source.

Meta launched new chips and reportedly started prepping for its biggest-ever workforce reduction.

Yann LeCun made his long-awaited return.

Perplexity tried to resurrect the personal computer.

And Google quietly shipped useful AI into your car, your Docs, and your Drive.

Welp. A lot happened. Here is everything you need to know. 👇

What do you do when your chip supplier decides to build the AI models competing with yours?

That is now the actual situation for OpenAI, Anthropic, and Google.

According to financial filings reported by Wired, NVIDIA has disclosed plans to invest $26 billion over the next five years to develop open-weight AI models.

That moves NVIDIA from pure hardware dominance into direct competition with the companies buying its chips.

The tension is structural. NVIDIA manufactures the hardware powering AI systems industry-wide, and its own models will almost certainly be optimized to run best on its own hardware.

AMD's CEO has already flagged open source approaches as essential to staying competitive, signaling how much this shakes the broader landscape.

What it means: 

There are three ways NVIDIA wins here. Closed-source labs pay more to keep up with NVIDIA's own models. Open source developers push the ecosystem forward. And consumers running models locally will likely buy NVIDIA hardware to do it.

Meta introduced four new in-house AI processors this past week, all part of its Meta Training and Inference Accelerator family, known as MTIA.

The lineup: MTIA 300, 400, 450, and 500.

These chips handle generative AI and recommendation model workloads, and they scale in server racks of up to 72 chips each — matching NVIDIA's own rack architecture.

Meta says the MTIA 400 is the first chip in the family to deliver both cost savings and raw performance competitive with leading commercial products, directly targeting NVIDIA and AMD.

The 450 and 500 build on that with faster speeds and higher memory for more demanding AI workloads.

According to Reuters, Meta has already started deploying some of these chips and plans broader rollout through 2026 and 2027.

This follows the same playbook Google, Amazon, and Microsoft already ran — build your own silicon to reduce dependency on NVIDIA. In 2026 alone, those four companies plan a combined $650 billion in capital expenditures, most of it aimed at AI infrastructure.

What it means: 

Building your own chips is now the default move for any company serious about AI at scale. Meta's pace is what stands out here — roughly one new chip generation every six months. That cadence signals urgency. Whether their actual AI models can keep up with the hardware investment is still the open question.

Yuuuuup. The same week Meta announced new chips and massive infrastructure commitments, Reuters reported the company is preparing for its largest workforce reduction in its history.

According to Reuters, Meta is considering cutting 20% or more of its workforce — which could affect over 15,000 employees.

No official date has been set, but top executives have already told senior managers to start planning how to operate with a smaller team.

CEO Mark Zuckerberg has simultaneously been offering AI talent compensation packages reportedly worth hundreds of millions of dollars over four years.

Humans out, infrastructure in.

The timing adds weight to the concern. Meta's Llama 4 models drew heavy criticism over misleading benchmark results. Its largest planned model was cancelled. And its next-generation model, codenamed Avocado, has now been pushed to at least May.

That is nearly a year since Meta's last meaningful model release.

What it means: 

Meta is cutting people while committing $600 billion to data centers by 2028, and still has no competitive frontier model to show for it. The Moltbook and Manus acquisitions are interesting. But spending at that scale without a model that can compete at the top level is a gap that keeps widening.

Microsoft launched Copilot Health this past week, a dedicated feature inside Copilot designed to help users understand their medical records, wearable data, and health decisions all in one place.

Copilot Health pulls together data from smartwatches, fitness rings, and medical records to offer personalized insights.

The tool was built with input from over 230 physicians across 24 countries and sources health information from Harvard Health, using the National Academy of Medicine's standards for credible sources.

Users can connect records from multiple providers through a platform called HealthEx and delete their data at any time.

Health data in Copilot Health stays separate from regular Copilot conversations and is not used to train AI models. It is also not HIPAA-protected.

It is technically not intended to diagnose anything. Let's be honest though — that is exactly what most people will use it for.

Currently waitlist only, launching for US adults in English.

What it means: 

OpenAI launched a health product. Anthropic went the plugin route. Now Microsoft is going all in. High-cost, knowledge-based professions — healthcare, accounting, consulting — are getting disrupted faster than most AI strategies account for. The models are reasoning faster and more accurately every month. That disruption timeline just got shorter.

One of the most prominent names in AI left Meta, raised $1 billion, and is betting the entire current AI paradigm took a wrong turn.

Yann LeCun, former chief AI scientist at Meta, has officially launched AMI — Advanced Machine Intelligence.

The Paris-based startup focuses on building what LeCun calls world models: AI systems that understand the physical environment the way humans and animals do, moving beyond text prediction entirely.

AMI launched with over $1 billion in seed funding, one of the largest early-stage AI investments in European history.

Investors include NVIDIA, Toyota, Samsung, former Google CEO Eric Schmidt, and Amazon founder Jeff Bezos.

What it means: 

LeCun has argued for years that large language models are sophisticated pattern matchers, not real intelligence.

Now he has a billion dollars to prove it. Whether or not the LLM ceiling is real, AMI is aimed squarely at where current AI still fails most — robotics, physical world tasks, and reliable real-world reasoning.

There is a war brewing over where your AI agent actually lives.

Perplexity launched Personal Computer this past week — an autonomous AI agent designed to run persistently on Mac devices and handle long-running tasks on your behalf, around the clock.

Unlike typical AI assistants waiting for prompts, Personal Computer operates continuously in the background on your local machine while connecting to Perplexity's hybrid cloud infrastructure online.

This builds on Perplexity Computer, the company's cloud-based autonomous agent system that already uses around 20 frontier models to complete tasks.

Personal Computer brings those capabilities to your actual hardware.

The context matters. OpenClaw has become arguably the most popular open source software right now, largely because people want agents running locally on their own machines. Perplexity is positioning Personal Computer as a more secure, more polished alternative.

Mac only, currently waitlist-only for Perplexity Max subscribers.

What it means: 

The personal computer is not dead, apparently. lolz. The OS just stopped taking instructions and started taking objectives. Perplexity marrying its impressive cloud agent infrastructure with local machine access is a smart wedge into a space OpenClaw currently owns. The community gap is real, but the product polish could matter more than people think.

Nah, this one barely got covered. And it really should have.

According to FedScoop, the US Senate has officially approved staff use of generative AI chatbots with official Senate data — a significant shift in government technology policy.

Senate staff can now use Microsoft Copilot, Google Gemini, and OpenAI's ChatGPT with official Senate data following approval from the Senate Sergeant at Arms and Chief Information Officer.

Each Senate employee is eligible for one license to either Gemini or ChatGPT at no cost.

What it means: 

Handing AI tools to undertrained, technically unfamiliar users at the legislative level is how AI slop finds its way into actual policy.

The average US senator is in their mid-60s. More than a third are 70 or older.

Recent congressional history includes senators who didn't understand how Facebook makes money, couldn't identify who makes iPhones, and once described the internet as a series of tubes.

Access without serious training investment is not a win. It's a liability. The federal government needs to treat AI education as urgently as it treats AI access.

Google announced this past week that Gemini is now integrating directly into Docs, Sheets, Slides, and Drive.

Users can now prompt Gemini to draft documents, build spreadsheets, or create slides by referencing specific emails, meeting notes, or files — pulling it all together without manual information gathering.

In Google Docs, Gemini can generate first drafts, rewrite sections to match a specific tone, and reformat documents around reference notes.

Sheets users can ask Gemini to build checklists and contact lists by pulling data directly from Gmail and Drive.

Google Drive search now features an AI overview, surfacing relevant files, pulling citations, and answering questions across documents, emails, and calendar entries.

A new prompt bar now appears directly inside each Workspace app.

Rollout starts in beta for Google AI Ultra and Pro subscribers, with Docs, Sheets, and Slides going global in English first and Drive features launching in the US initially.

What it means: 

Google has been shipping useful AI everywhere, quietly and consistently. From Maps to your car to your full Workspace suite, Gemini keeps showing up where people actually work. This is not a flashy announcement. It is just genuinely useful. And in enterprise AI adoption, consistent and useful beats flashy every single time.