Welcome back! This week, we have very exciting news across hardware and software, with the highlight being the realization that recursive self-improvement, where AIs create or improve other AIs as the primary driver of progress, is here.

Beyond that, we discuss NVIDIA’s latest hardware, new frontier models, and new products, all that you need to know to stay in line with AI progress.

Enjoy!

This week, Jensen Huang delivered his big GTC 2026 keynote, one of the key events for learning NVIDIA’s roadmap for the future and, by extension, of AI.

And as I predicted, besides the already viral $1 trillion in revenue guidance for 2026 and 2027 (that only includes Vera and Grace GPU platform sales), which doesn’t include other components like CPU demand, the most important part of the keynote was the announcement of its first disaggregated inference hardware platform, combining its own GPUs with Groq’s LPUs, fruit of their acquisition of the latter a few months ago for $20 billion.

But what is a disaggregated hardware stack? AI workloads are not uniform; they vary a lot depending on what phase of the process you’re in. Simply put, GPUs are not the optimal solution in some cases.

Thus, NVIDIA has gone ahead of the market and prevented this from becoming a real problem by acquiring one of those solutions that are superior to GPUs in some aspects of AI inference and by building a server that combines GPUs with the new hardware.

The promise? 700 million generated tokens… per second.

So… is this NVIDIA’s new golden goose?

In principle, this can represent an important narrative shift regarding NVIDIA; they seem to be clearly transitioning out of “general-purpose land” and progressively building ad hoc/ASIC hardware.

And to be clear, the move seems logical and inevitable, given that disaggregated inference is widely believed to be the future of AI inference infrastructure.

However, do we actually need ad hoc hardware for this? My gut says yes, but here’s the thing: software matters a lot too.

Because the inevitable thing to question here is: what about AMD? AMD has not announced a disaggregated inference hardware stack, so they must be tremendously behind, correct?

Well, actually, no. If we look at the performance of both AMD and NVIDIA GPUs on disaggregated serving, AMD is consistently cheaper at the same performance level (serving 100 tokens per second to every user) as NVIDIA, and the gap is actually widening.

The comparison between NVIDIA and AMD is now as noisy as ever, for three reasons:

I’m not ready to make any bold predictions, but I’m a shareholder of AMD, and it might be the case that this stock is savagely underestimated. How will we know if that’s true? Once the MI455 servers start hitting the market this year.

Elon Musk has announced that Tesla is becoming a semiconductor fabrication company, with the launch this weekend or at the latest on Monday.

The argument is that Tesla wants its own semiconductor manufacturing capacity because outside suppliers (e.g., TSMC, Samsung, or Intel) may not be able to provide enough advanced chips for the company’s future AI, autonomous driving, and robotics ambitions.

In practical terms, the thesis is that Tesla is trying to remove a future bottleneck. Musk has said chip supply from partners such as TSMC and Samsung may not be sufficient for the scale Tesla wants, and that the company may need a much larger, domestic fab with logic, memory, and packaging capabilities.

To be brutally honest, this is a gargantuan effort you have every reason to be skeptical of.

If anything, this only makes sense to me if their idea is to manufacture Tesla’s chips on older process nodes. The idea is that AI is mostly memory-centric (especially for inference), meaning that the real bottleneck is not compute but memory. Thus, it isn’t necessarily required to build your chips using the most advanced technology as NVIDIA does in order to remain attractive to customers, chips that require Taiwanese or South Korean efforts.

Instead, if Musk envisions this as mostly for internal demand, they can get away with using much older process nodes that require fewer investment, and with enough know-how in the US to accomplish what is still a gargantuan task.

However, from an investor’s perspective, this makes it even harder to delucidate what ‘Tesla actually is’.

A car company? A battery company? humanoid company? A chip manufacturing company? At this point, I think even Elon would struggle to explain what an investor is buying when buying Tesla.

In my humble opinion, all roads lead to a potential merger between Tesla and SpaceX, the equivalent of a ‘superapp’ in companies; a so-called “engineering company”. That narrative could be sexy enough to get me on board.

But for now, I’ll stay on the sidelines because I simply don’t understand what this company is anymore.

Stripe is not going to let the agentic revolution escape. They have just announced the MPP protocol, Machine Paymanets Protocol: an open standard for machine-to-machine payments, so AI agents and services can negotiate and complete payments programmatically.

In plain terms, it lets an agent encounter a priced resource, pay for that specific action or request, and continue the workflow without dropping into a human checkout flow.

The problem it solves is that today’s payment stack was built for humans, not autonomous software. A typical purchase often requires creating an account, navigating pricing pages, choosing a plan, entering payment details, and setting up billing, which pulls a human back into the loop and makes fine-grained agent payments awkward or impossible.

MPP is meant to remove that friction by providing agents and services with a standardized way to coordinate payments for per-call usage, microtransactions, and recurring charges.

In other words, if you ask your agent to buy something on Amazon for you, and Amazon sets up MPP, the agent can programmatically search for your item and immediately pay for it on your behalf, without you having to set up a credit card or API token for your agent. It will just work.

The growing belief in Silicon Valley is that, in the future, most Internet participants will be AIs. Thus, the natural progression is to treat agents as first-class citizens of the Internet.

Would I trust an agent with my money today? Hell no, but we’ll eventually get there, and probably sooner than later.

A new trend is brewing in the AI space: recursive self-improvement. AI Labs and researchers talk big game about their own AIs updating themselves to get better.

But how does that look like?

Well, Andrej Karpathy’s autoresearch repo might be the answer. The idea is pretty simple: give an AI agent a training setting and a compute allowance, and let it run overnight to improve training results.

And as Karpathy showed in the thumbnail, it works. Agents consistently run experiments, try new stuff, and make progress toward lower and lower error rates.

For example, Shopify’s CEO, Toby Lutke, used this small project to improve performance on a templating engine that had remained untouched for 20 years by 53%.

Recursive self-improvement is what many believe truly kickstarts the acceleration. Once AIs can improve AIs, human imagination and time are no longer the bottleneck; the bottleneck is compute.

Again, investors remain somewhat convinced that AI is in a bubble and unwilling to go all-in on the industry, but chances are we'll realize quite soon that we don’t have nearly as much compute as we think.

Nonetheless, GPU rental prices, to the dismay of AI doomers, are rising rapidly even for older GPUs.

Minimax has released Minimax M2.7, and it seems like they have hit a home run. The company, which recently IPOed, has shown remarkable progress with its models (or rather, systems), even though it was one of the companies directly accused by Anthropic of distilling its models.

The biggest highlight comes from its efficiency. As you can see below, Artificial Analysis places them at the top of the Pareto efficiency board, offering the best performance-to-cost ratio amongst both open and closed models.

But what makes it worth talking about is that the majority of the improvement came not from the model but from the harness.

As they explain, they used the model itself to self-improve (another hint to the story above). The model could review failed attempts, suggest changes, update parts of its workflow, test the results, and keep improvements that worked.

The improvements it describes are mostly about workflow: better settings, clearer instructions, better loop control, and better ways to handle memory and repeated mistakes. In some cases, this system improved performance by up to 30% relative to an unoptimized baseline.

The elephant in the room for US Labs is not whether Chinese Labs will create better models. That seems extremely unlikely, given how underserved they are in terms of compute and the fact that compute is the main driver of progress.

But where ‘Chinese AI’ can do serious harm is in the enterprise. Nothing matters more than performance-per-dollar in enterprise settings, so companies will naturally be tempted to adopt Chinese models (which are free beyond the compute required to serve them) and are ironically more secure because you can just download and store them in your organization, without having to worry about data leaks.

People don’t seem to notice this yet because most enterprises aren’t sophisticated enough to run open models, but, in my view, it’s undoubtedly the future.

Anthropic has conducted an 81,000-person study to learn what people feel about AI, and it’s quite eye-opening.

The main limitation is that this was not a representative public survey. It was based on people already using Claude, so the results likely overstate how positive the general population is about AI.

Great study; more such data is needed to truly grasp the general sentiment regarding this technology.

Of course, we need to take results with a pinch of salt; the sample distribution is already predisposed to being AI-favorable (a lot of people in our society blatantly refuse to use AI, and these aren’t represented here), but it’s still a good exercise for AI Labs to understand how their users feel about them.

Anthropic has announced its first 1-million-context window, more than a year after Google. But for such an important event, it’s never too late.

Crucially, and unlike Google or OpenAI, there is no price increase above a certain threshold, keeping the token price flat throughout the entire context at the usual $5/$25 per million input/output tokens.

Also, the model seems to be strongest against ‘context rot’. One thing is to claim a one-million context window; another is to maintain quality even at the later ends of the context.

Known as ‘context rot’, models often see severe performance declines as they are incapable of picking up long-range patterns or identifying ‘needles’ (low-frequency facts or data present in the context but not repeated, making them harder to find). And for this, Claude seems to be the best at preventing it.

It’s surprising they kept the pricing the same across the entire context. To be clear, they are still the most expensive solution by far, which suggests they might just be subsidizing the marginal cost, but maybe they’ve found a better attention solution.

The problem with long context is two-fold:

So a model with superior context visibility (the opposite of rot), coupled with constant pricing, may suggest something more than just a context window increase: an algorithmic improvement. As a matter of fact, we already know attention algorithms, like DeepSeek Sparse Attention, that offer ~linear scaling for attention, so this might be the real reason behind the improvement.

At GTC 2026, NVIDIA announced Vera Rubin Space-1, a computing platform designed for orbital AI data centers.

NVIDIA says the system is designed for space missions and constrained environments, and that it includes components such as IGX Thor and Jetson Orin.

CEO Jensen Huang framed it as part of a push to move AI compute closer to where space-generated data is produced.

Interestingly, Jensen himself addressed the elephant in the room: the cooling problem. Heat doesn’t dissipate in a vacuum. The core problem with space-based data centers is that you lose the most effective cooling mechanism used on Earth: convection.

On Earth, data centers rely heavily on air or liquid cooling, in which heat is carried away by a moving fluid (air or water). In space, there is no atmosphere, so:

Radiative cooling is fundamentally much less efficient. He acknowledged this is an unsolved problem, but hey, maybe they manage to pull it off.

Mistral has announced the release of Forge. Mistral is introducing it to help enterprises build custom, frontier-grade models on their own internal data, rather than relying solely on general-purpose models trained mostly on public data.

The pitch is that companies should be able to train models that actually understand their proprietary codebases, policies, workflows, and domain language.

It is also clearly built with agents in mind. The article argues that custom-trained models make enterprise agents more reliable because they can follow internal procedures, select tools more accurately, and handle multi-step workflows in ways that align with how the organization actually works.

Technically, Forge supports several stages of the model lifecycle, including pre-training, post-training, reinforcement learning, evaluation, and both dense and mixture-of-experts architectures, with multimodal support where needed. The broader message is that Mistral wants enterprises to treat models less like off-the-shelf software and more like long-term strategic assets that can be continuously improved.

Despite being European myself, I’m a professional ‘Europe AI’ hater; Europe’s AI trajectory is plagued with errors. But this isn’t one.

I’m a profound believer that fine-tuning, training models on new data, is the quintessential unlocker of enterprise AI. The rational feels almost stupid to explain; AIs are software trained on data. The more data they see on a given task, the better they work.

Is it so revolutionary to think that, maybe, what enterprises are missing is just to train the AIs on their data?

Nonetheless, Thinking Machines Labs, the AI lab packed with star researchers from OpenAI, Meta, Anthropic, and DeepMind, is structuring its entire value proposition around this idea, offering fine-tuning-as-a-service. This is so clearly the future of this industry, I can’t even hide my excitement.

OpenAI has announced the release of GPT-5.4 mini and nano, two smaller models in the 5.4 family with very competitive pricing.

GPT-5.4 mini is a smaller, faster version of GPT-5.4 (i.e., it’s a model distillation), optimized for coding, computer use, multimodal understanding, and subagent workflows, available across all channels (ChatGPT, Codex, APIs).

Availability differs by product. In ChatGPT, Free and Go users can access it through the Thinking option, while other users get it as a fallback after reaching GPT-5.4 Thinking limits.

In Codex, it uses about 30% of GPT-5.4 quota, making simpler coding tasks cheaper. API pricing is $0.75 per 1 million input tokens and $4.50 per 1 million output tokens, placing it in a very competitive zone but raising questions about comparisons with Chinese models, which are even cheaper and, with the size discount, may be closer in performance than one would think.

It’s clear that OpenAI has shifted its strategy toward enterprise use cases (coding, maths, STEM in general) and away from more conversational approaches, which now feel compressed into a single model: GPT-5.3 Instant.

This pivot has already been discussed in mainstream media, and it’s totally true that the latest models feel much more, for lack of a better term, “emotionless.”

I applaud this move, as I fear overly sycophantic models can truly make people lose their minds (LLM psychosis is a very real thing these days), but it will come at a cost; more and more users are moving to other platforms with more welcoming models.

Additionally, they’ve released subagents in Codex, meaning you can now spin off several agents in parallel to work on a common task.

I really don’t have much to say here (no benchmarks, no nothing to comment on), but in case you’re wondering, my general recommendation about multi-agent settings is to avoid them unless they explicitly clarify that the AIs were trained to work as a group, something Moonshot did with their agent swarms. Without that joint training, subagents are usually just a waste of money.

Anthropic has released a cool feature called ‘Remote Control’ that basically SSHs (remotely connects) your mobile Claude to your local Cowork instance and lets you continue communicating with your agent even when you’re outside.

A truly declarative paradigm, voice-to-work, is coming. The ability to have your personal AI secretaries handle your business while you’re walking the dog is extremely appealing.

Yes, we have to remind ourselves of the reliability issues, but I’m pretty confident we’ll get there; there’s too much value in achieving this use case not to have it work eventually.

Lovable has pivoted from creating apps to a general assistant (i.e., the 3829339 start-up doing an OpenClaw/Cowork pivot). The launch video shows different examples of Lovable creating pitch decks, reports, and a meal-planner app.

But let me explain why this is a really bad sign.

Tremendously bearish. Such a pivot for a company that proclaimed itself the fastest-growing company not only in Europe but in history, allegedly being the fastest to $100 million in annualized revenue (and then allegedly $400 million as of March 2026), makes literally zero sense, if not for one of two things:

I’ve talked about this in the past. Lovable is going to fail, and it’s going to be one of the biggest failures in the industry. And to be clear, I’m biased. I don’t like them.

As mentioned earlier, I’m European, so naturally I cherish (and envy) having the US’s AI ecosystem here in Europe.

That train looks largely lost at this point, but it sure doesn’t help having one of its primary hopes, Lovable, being so sketchy about numbers and, most importantly, blatantly lying about their “European status” as well; they are registered in Delaware, which, if my geography doesn’t fool me, is a US state, not a European country.

But I digress. This move only makes my suspicion worse, as pivoting from an allegedly money-printing machine to the most competitive industry in AI today is simply ‘Pleasing nervous investors 101’ and one of the first signs things are turning south for them.

But I believe the problem goes deeper. I’ve insisted on this a lot; no AI startup survives without its own models.

If most of your business is literally an API call away, you don’t have a business; you have a few prompts, a frontend, and a hopefully talented marketing team all pretending to be one. You can pretend to play the AI game for a while, but without ‘owning the AI’, your efforts look totally unserious and soon obsolete.

Cursor got the memo and now has proprietary AI that is actually great (more on this below). Lovable didn’t, and I wonder whether the fact that its founders are ex-McKinseys rather than actual engineers has anything to do with this tremendous lack of vision.

Talking about Cursor, it has released Composer 2, the new version of its coding agent.

The core claim is that Composer 2 improves substantially over earlier Composer versions on the benchmarks. According to them, it’s basically a frontier-level coding model.

It reports 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual, some of the most popular coding and agents benchmarks, showing performance that can actually topple the frontier models in some cases (see thumbnail).

Cursor attributes the improvement to two training changes: its first continued pretraining run, which it says created a stronger base model, and reinforcement learning on long-horizon coding tasks.

In plain English, they haven’t trained an entirely new model; instead, they've used an open-source model (Moonshot’s Kimi K2.5) and run additional training runs on top.

A highlight is the pricing. The standard Composer 2 model is listed at $0.50 per million input tokens and $2.50 per million output tokens. For reference, that is ten times cheaper than Anthropic’s top model, Opus 4.6.

Cursor also says there is a faster variant with the same intelligence priced at $1.50 per million input tokens and $7.50 per million output tokens, and that this fast variant is becoming the default.

In case you’re wondering, this ‘fast’ model is not a different model; it’s just run on smaller batches (i.e., your requests are given higher priority and processed more quickly). They didn’t mention that, but they ain’t fooling me, I do this for a living.

But how do I know that is the case? The key reason I know this mostly about batching is that they mention “same intelligence”; in AI, you can’t increase both speed and intelligence, it’s either one or the other.

The only possible way is if Composer 2 was a dense model and this new fast one was a mixture-of-experts, but I can guarantee you that both are already MoEs. Another option would be the fast version using MTP (Multi-token Prediction), which outputs multiple words rather than just one, but that would impose a performance tax.

Therefore, the only option left is providing faster responses with more aggressive batching. This is super expensive to do, thus the considerable price rise.

This is the way to survive the Big AI Lab tyranny: progressively using your new data to train your own models and slowly but steadily migrating away from third-party models.

I believe this conversation doesn’t get remotely discussed enough. Big AI Labs are priming all of us to become ‘hooked’ on their models while heavily subsidizing subscription and API prices.

Nevertheless, Cursor is already eating up huge losses per subscription (some rumors put the number as high as $5,000 inference costs per each $200 Ultra subscription), so it’s crucial for them to have powerful in-house models that don’t carry the hefty Anthropic margin, in case they expect to make money someday, which I assume is the case.

As I told a client yesterday, there’s a false belief that AI is cheap. It’s not; it’s not only not cheap today, but once AI Labs start to charge the real prices, it’s going to be pretty terrifying.

In a release I can’t help but be skeptical about, the start-up Okara has released the first-ever “AI CMO”. You provide your company’s website, and the system deploys a series of agents to optimize your web presence.

I completely see the value this could bring; I just feel like it’s mostly going to be slop.

They provide zero proof on fine-tuning. Is this a simple prompt-engineered ChatGPT, or an actual growth-optimized model? If it’s the former, this is just a slop machine; ChatGPT is not better at copy than a human. In fact, if you look at the video, all you see is AI slop of the purest kind.

There’s promise in the idea, but I’m highly skeptical of execution and also how defensible this business actually is: with no proprietary data or models, what stops me from building the exact same system for myself?

Google just redesigned both AI Studio and Stitch around a similar goal: making AI tools feel less like one-off generators and more like places to actually build.

The new Google AI Studio redesign pushes the product closer to a full app-building environment. Instead of mainly being a place to test models and prompts, it now looks more like a workspace for creating real applications, with stronger support for coding, iteration, and deployment. Hype aside, it’s mostly yet another agentic coding app like Codex or Claude Code.

The Stitch redesign does something similar for design. It moves beyond simple screen generation and becomes more of an AI-native canvas for creating, refining, and exploring UI ideas. The emphasis is on faster iteration, more flexible workflows, and a smoother path from concept to prototype.

You can tell Google feels the coding market slipping away and is trying to jump back on board while it can.

In particular, they are super strong on the creative side; Stitch is rapidly becoming the new king of AI-led app design and creation.

However, what they are missing is the building part; the experience on Codex or Claude Code seems to be much more appealing today, probably mostly because GPT-5.4 and Opus 4.6 are head and shoulders better than Gemini 3.1 Pro at coding (or, at least, that's what seems to be what most people believe).

I’m a massive Google bull, but at least today they seem a little behind OpenAI and Anthropic in agents and coding.