Welcome back! This week, we have a long list of interesting news to talk about. From China’s latest SOTA model to the US’s next great small model, we also tackle NVIDIA and Microsoft super events, Uber clamping down on AI costs, Bernie Sanders’ latest AI rant, and many more.

Enjoy!

Minimax has released a new model that, on paper, is competitive with the best the US has to offer. Across several benchmarks, it holds its own against Opus 4.7 or GPT-5.5, alongside Opus 4.8, the bleeding edge of the industry.

Despite being considerably smaller than its rivals (at least ten times smaller), it competes and offers a massive one-million-context window, which suggests this model punches well above its weight class.

I wrote about it in way more detail here, but the crucial thing to highlight is that, much like all other Chinese Large Language Models (LLMs), they perform some sort of compression over context.

This stems from the fact that, while you might not remember what you had for dinner three days ago, you can recall decades-old childhood memories with ease, because the human brain is opinionated about what’s “worth remembering”.

However, this is extremely complicated to do with AIs, to the point that we largely avoid doing so if we have the compute means to avoid it.

This results in models that do not compress context; if you send them an entire 50-page report, they’ll store in context every single word in it, which means that an LLM’s context grows proportional to its length (and in a quadratic fashion, actually; tripling sequence length nine-folds the computational requirements).

Chinese Labs, much more compute-constrained, are forced to innovate in this regard, and they do so by forcing this context compression.

But how?

Say you’re reading a 100,000-word book and you want to predict what will happen in the next few hundred words. To do so, you’ll probably bear in mind what has recently occurred, while also taking into account a summarized view of what happened in earlier chapters. You don’t remember every single thing that happened, only those things your brain considers important.

But if we give an AI a 100,000-word sequence to predict the 100,001st word, the model will store all previous 100,000 words. Not a single one is ignored. Even if a word is ‘uhm’ or ‘ehhh’, they are also attended to.

This is as if, for you to decide what to eat today, you considered not only what you ate recently, but also what you ate a year ago from today, while also taking into consideration yesterday’s debate with your husband about which cushion color works best on your sofa.

This seems dumb, but it’s exactly what is going on; all options are considered.

To handle this, Minimax proposes a “hardware-aware” compression. The context gets considerably compressed, in the same way other Chinese models like DeepSeek V4 do, and it’s also designed to benefit GPU architectures the most, basically by ensuring that, for every retrieved data bundle, as much of it as possible is ‘useful.’

This is because, given the nature of DRAM, the memory used by GPUs, it’s faster to retrieve two contiguous data points than two separate ones: the first two can be retrieved directly, whereas the latter requires two retrievals (retrievals are fast, but they still add up to delay if done suboptimally).

This is a marvel of engineering. But is it really SOTA?

The jury is out, but the answer is almost definitely no. There’s no such thing as a free lunch, and the only reason Chinese Labs are running these compression mechanisms is that they're being forced to.

US labs don’t run these mechanisms because they have the compute and the capital to avoid doing so. The rationale is simple: if you force the model to compress, something that might be needed is forgotten, which obviously affects performance.

Chinese models are smaller, too, making it even harder to believe, because I’m afraid not, there’s no such thing as a Chinese secret sauce for now.

But Chinese models are reaching a level of performance-per-dollar where they become ‘best value’ alternatives that could seriously undermine OpenAI's and Anthropic's ability to penetrate enterprise budgets.

A couple of days ago, Jensen Huang did a keynote at Computex, Taiwan’s annual expo for the semiconductor industry, one of the most important events of the year. And Jensen had several things to announce.

While the obvious highlight was Vera Rubin's servers being deployed finally, to me, this was not the key takeaway.

The key insight to highlight from Jensen’s keynote was a claim Jensen made: he projected that CapEx per GW would grow from $50 billion today to around $100 billion soon, clarifying that AI hardware is not only not becoming cheaper, but it’s also more expensive than ever.

How companies expect to make money from this business, I truly don’t know. I always say that I don’t believe Anthropic or OpenAI can be profitable as long as NVIDIA, Hynix, and the semiconductor players are commanding huge gross margins.

While AI software is commoditized, or at least there are several players competing, which forces prices to fall, these same companies are paying larger premiums than ever to access the hardware intended to train and run these models.

Jensen would push back, saying that tokens/watts are falling, so every GW allows for much larger revenues, but customers around the world are already struggling with token costs today. So either our ability to generate tokens increases by several orders of magnitude per dollar, or this business won’t make sense for the foreseeable future.

As Google shows below in the market section, AI will require substantial liquidity to survive over the next few years, including the inevitable participation of all of us, willing or not.

Because if you think you have a choice, you don’t, because OpenAI and Anthropic are getting into your favorite index funds, and fast (Nasdaq has reduced the required time as a public company to get inserted from one year to 15 days ahead of the SpaceX IPO).

And even if you don’t own index funds, they'll still get into your 401(k)s through index funds. Directly or indirectly, you’re going to be an owner of these companies whether you like it or not.

And listen, “they”, and I mean the powers that be, be that the US, Germany, or Spain, don’t have a choice; they need to make all of us part of their big bets to sustain them, in the same way they are going to push crypto stablecoins down our throats eventually, too, in order to distribute this unpayable debt that the largest economies in the world have amassed.

And even if you somehow avoid all of that, you’re still going to pay with inflation, because the combination of trillions of AI private credit and government debt has to be paid, and the only way is to deflate the value of our currencies.

Yesterday, Microsoft had its Build 2026 event, and I must say I am happy with what I saw.

Microsoft announced MAI-Thinking-1, its first in-house reasoning model, described as a 35-billion-parameter model for multi-step reasoning, long-context work, and code generation. It also introduced MAI-Image-2.5, MAI-Transcribe-1.5, MAI-Voice-2, and MAI-Code-1-Flash for GitHub Copilot and VS Code, up to 7 highly competitive models from scratch.

This shows that, finally, the acqui-hire of Inflection two years ago is starting to bear fruit. For instance, their image model now ranks second only to GPT-image 2 in a popular image-editing benchmark (thumbnail).

Agents were also a highlight. Microsoft introduced Microsoft Scout, an always-on personal work agent built on OpenClaw and Work IQ, designed to operate across tools such as Teams, Outlook, OneDrive, and SharePoint. You know my skepticism with current agents, so we’ll see how that goes.

Microsoft also previewed Project Solara, described as a chip-to-cloud platform for an “agent-first” computing model, and announced Surface RTX Spark Dev Box, a local AI development machine expected later in 2026 in the US, pending authorization.

And on quantum computing, Microsoft introduced Majorana 2, saying its qubits are 1,000 times more reliable than the previous generation and that the company is targeting a commercially relevant quantum computer by 2029.

The paper they released on MA1 was super impressive; a gold mine of research nuggets that tells me Microsoft AI is finally on the right track in terms of their AI efforts. The model was clearly not designed to score highly; it’s a Sonnet 4.6-level model with “just” one trillion parameters, but the process they used to train it signals sophistication and shows us a company that is finally serious about its internal AI efforts.

And best of all, committed to open-source too.

To highlight a particular value, I was shocked by their super-low MFU of 20% despite being a training run; this means GPUs were running at 20% of peak theoretical compute—it does not mean they used only 20% of the 8,000-strong cluster.

This peak is unattainable, but such a low score, considering that the Llama 3 team in 2023 had around 40%.

Unsurprisingly, the stock still fell after the event, quite a bit actually, which is what it has gotten all of us used to lately. However, in my opinion, that’s just short-termism by investors; if anything, Microsoft improved in my eyes, even if the outcomes of these efforts will take some time to show up in the P&L.

But before we move on, I really have to mention two key metrics: one that Mustafa Suleyman mentioned, another that he disclosed without intending to (at least, on paper).

For their MAI-Thinking 1 model, they claim that their Maia inference system (their new inference chip) offers 1.4x better performance/dollar relative to NVIDIA’s GB200 server.

That is quite the claim, and I’m sure Jensen is going to call them out for it. But if true, NVIDIA investors should be worried unless Vera Rubin blows everyone and everybody out of the water. We’ll see.

The other one merits a round of applause for me, because I quite literally nailed Mythos’ training budget, at around 2×10²⁷ FLOPs, as I predicted in this Medium article a week ago, thanks to the fact Microsoft, well, told us:

But unless you’re an AI geek like me, that number probably doesn’t say much beyond the fact that it represents a number with 27 zeros, which suggests that it’s a lot.

But how much?

This means that you could train 100 GPT-4s with the amount of compute that was used to train Mythos.

That is the scale of AI today, and as we’re about to see with the Vera Rubin piece below, it’s only going to get crazier, which raises the question: Is training on a gazillion tokens really the only way we have to improve AIs?

Uber has limited its per-user AI spending to $1,500 after blowing past its annual budget in less than four months. It’s not surprising, given that both its CTO and COO voiced concerns about the ludicrous spending one can rack up with AI.

In particular, the former even added that they were struggling to see returns on such spending, making this budget constraint a matter of time and leading to today’s decision.

People may panic about this, but this is just normal. Expected. I got criticized a lot when I said ‘tokenmaxxing’ was a stupid strategy, as if generating more tokens would magically transform businesses.

Yet the only meaningful transformation so far is your OpEx. Robert Solow once said, “PC are everywhere, but in the statistics.”

I now propose the AI version:

As published by Bernie Sanders on X, the senator said he will introduce a bill to give the public a 50% ownership stake in the largest AI companies in America. He said the goal is to ensure that wealth created by AI is used broadly and to give the public the power to block company decisions that could harm Americans.

Sanders’ Senate site identifies the proposal as the American AI Sovereign Wealth Fund Act. It would create a sovereign wealth fund through a one-time 50% tax paid in stock, not profits, from major AI companies such as OpenAI, Anthropic, and xAI (SpaceX).

Under the proposal, the federal government would receive voting shares and equal board representation at covered companies. Sanders argues the fund could grow with the value of AI firms and eventually support direct public benefits or public programs.

I’m all for finding ways to redistribute the benefits of AI across society. I also entertain the idea that these companies stole our data to train their models and that they are, in a way, in debt to society.

But fighting theft with theft, which is what this is, is not the solution.

Even if I’m not necessarily super pro-taxes (I pay an eye-watering amount of taxes in Spain relative to what I earn, basically crossing the confiscation barrier at this point), I understand their value and believe they make much more sense in this case: tax positive outcomes; don’t intervene in the search for justice.

And to be clear, I do think AI incumbents have to be very careful about inequality of outcomes, or they are going to suffer massive social rejection. Maybe UBI (Universal Basic Income) could be another option, too.

Something will have to be done eventually if AI is so transformational. But that something should not be theft.

Would I support a purely economic intervention? That is, having the US taxpayers buy a 50% stake by paying $500 billion to OpenAI or Anthropic?

Hell no. That is a bailout these companies have not earned the right to.

I don’t think taxpayers would agree to invest in companies that are so massively in debt, either. And I don’t think they should.

But once these companies have positive cash flows, maybe we could discuss how to ensure societal benefits. But the solution can’t be theft, which is precisely what Bernie wants.

As a European, what Bernie says sounds super reminiscent of what European politicians say. And that’s not a good thing. Far from it.

Please don’t let US lawmakers make the same mistakes that have condemned Europe to irrelevance and ostracism. Don’t let politicians ruin the US as they did with Germany’s industrial base, or with Europe’s energy security, just because an RBMK reactor with a positive void coefficient (which doesn’t exist anymore) and three irresponsible Russian engineers led to the Chernobyl disaster 40 years ago.

European politicians understood regulation and interventionism as the way to progress and wealth redistribution, and all they have achieved is a dying continent at the mercy of the US and China. I fear the US could fall into the same mistake.

I travel a lot to the US, and I see the same ideas that destroyed Europe’s future being thrown around too lightly. And yes, there’s a very real issue with inequality in the US, with the top 1% owns 37% of income and the bottom 50% owns only 2.5% of wealth.

That is not sustainable. But theft is never the answer.

And the funniest thing of them all: has AI actually proven to be that incredibly unequalizing force? Something that, in the words of Senator Sanders, “could become smarter than us and function independently of our control” and thus warrant intervention?

As I’ve reiterated countless times, this is just doomer porn. But just like I don’t think the US Government should steal 50% of a company from its owners, wouldn’t it be funny if it happened to the same people who pushed the doomer narrative in the first place?

Quite possibly the news of the week, Anthropic has filed its confidential S-1 for review to reserve its right to go public.

The company said it has confidentially submitted a draft Form S-1 registration statement to the US Securities and Exchange Commission for a proposed IPO of its common stock.

The filing gives Anthropic the option to go public after the SEC completes its review, but the company said the offering remains subject to market conditions and other factors.

Anthropic did not disclose how many shares it may offer or the expected price range. The announcement was made under Rule 135 of the Securities Act, meaning it is not an offer to sell securities or a solicitation to buy them.

Well, as a matter of fact, the company didn’t disclose a single thing. Extremely secretive despite having an alleged booming business “that will be profitable this quarter”, right?

This industry is just so full of shit it’s actually amusing to me.

For what it’s worth, it’s a great company, nobody doubts that. But just like with SpaceX, the problem will be its valuation (and surely the same will apply to OpenAI).

For context, they have just closed a funding round at $965 billion, which means it has to IPO way higher than that, easily above $1.5 trillion, which would make it as valuable as Meta, and maybe even closer to $2 trillion. Make that make sense.

If that’s the case, that’s hilariously overpriced, but I do think some investors will be blind enough to purchase it at that price. We’ll see.

As reported by Reuters, Alphabet, Google's parent, plans to raise $85 billion in equity to fund the expansion of its AI infrastructure.

Alphabet says the money will support AI-related computing capacity, data centers, and custom chip development. The move follows a sharp increase in planned capital spending, with Alphabet’s 2026 capex now guided at $180 billion to $190 billion, up $5 billion from the previous estimate. Alphabet shares fell after the announcement, and the gap with NVIDIA, the most valuable company on the planet, is now almost a trillion, up from “only” $200 billion a few weeks ago.

Interestingly, these $85 billion add to the $85 billion they’ve already borrowed over the last year.

A few weeks ago, I showed a graph showing that Hyperscaler FCF (Free Cash Flow, the amount of cash they have available for discretionary spending) was down sharply, meaning these companies were literally running out of money.

Google’s $85 billion raise clearly indicates that they are “all in” on AI and will do whatever it takes to win, even at the expense of shareholders like me.

Probably the highlight here is the participation of Berkshire Hathaway, which seems to have made up its mind on which AI horse they are betting on amongst the Hyperscalers (I wouldn’t count Apple as an AI bet just yet).

This is like IPOing again, and $85 billion less that could have gone to Anthropic and OpenAI.

Do the public markets have the $500 billion we’ll need to provide to Google, SpaceX, Anthropic, and OpenAI? I don’t think people realize how uncertain this question is.

As mentioned above, Jensen Huang has confirmed that both Microsoft and Dell/CoreWeave have deployed their first Vera Rubin NVL72 servers, the next-generation AI chips.

And the differences in performance, especially relative to memory and memory bandwidth (the key bottlenecks in inference, which is the primary driver of compute) to previous generations, are simply astonishing.

For example, in Rubin, each GPU inside a server can communicate more data to one another than a Hopper GPU could within its own package. This significantly increases the amount of data GPUs can share, thereby considerably elevating performance.

Think of it this way. In inference, you are bottlenecked by how much data can be moved. This means that your GPUs are “waiting” idle for data to arrive—they aren’t really idle, but running at very low capacity relative to what they could be doing. Therefore, by increasing the amount of data that can be shared, you’re significantly alleviating the bottleneck, leading to more tokens per second and per watt.

The one thing I need to point out is that some people are saying that in Vera Rubin, the speed between GPUs is faster than inside the package of an H100. But speed here is the wrong word; it’s bandwidth.

GPU-to-HBM speed, the time it takes a GPU to read data from its HBM chips, is on the order of hundreds of nanoseconds, easily an order of magnitude faster than the time it takes to reach other GPUs.

People confuse bandwidth (how much data can be sent per second) with speed (how fast it moves) all the time. Don’t be that person.

Google has introduced Gemma 4 12B, a mid-sized open model designed to run multimodal AI locally on laptops. The model handles text, vision, and native audio inputs, and Google says it can run on consumer hardware with 16GB of VRAM or unified memory.

Google describes Gemma 4 12B as a bridge between its smaller, edge-focused E4B model and its larger 26B Mixture-of-Experts model, with benchmark performance “nearing” that of the 26B model while using less than half the memory footprint.

Besides looking like a great model for its size, a key technical change is its encoder-free multimodal architecture: instead of separate encoders for images and audio, visual and audio inputs are integrated more directly into the language model backbone. For audio, Google says it removed the audio encoder and projects raw audio into the same space as text tokens.

This is genuinely interesting and something that I believe will become much more common with smaller models.

OpenAI “Sites” is part of the new Codex release, announced today. OpenAI describes Sites as a preview feature that lets Codex create and share interactive, hosted websites and apps.

Sites can turn ideas, analysis, and plans into dashboards, planners, review workspaces, project boards, galleries, and lightweight tools. The sites can be shared with anyone in a user’s workspace through a URL.

OpenAI said Sites is rolling out in preview for Business and Enterprise teams via the Codex app, and enterprise admins can enable it in admin settings. The company also said it is working with early partners, including Wix, Base44, Replit, Lovable, Figma, Webflow, and Emergent, as it builds a Sites partner ecosystem.

AI is innovating once again with yet another AI site generator. I swear this entire industry is three products being reinvented again and again.

What I will say is that the sites look incredibly clean, something OpenAI’s models have been really bad at historically. The reason seems to be its partner ecosystem, which might have helped improve such skills.

Rants aside, OpenAI seems to be looking to turn Codex into an enterprise platform, not just a tool to code, but a tool to build anything you want. For now, that probably pulls it closer to a slop machine than something actually useful; AIs still make a lot of mistakes.

And to be fair, this reminds me all too much of Confluence, the Atlassian tool that lets teams have sites, projects have repos, and all that stuff that nobody can justify its value, but everyone pretends it does.

Not trying to be overly cynical here, but this feels extremely tailored to the status quo in corporate America rather than something that helps it progress. Take this with a pinch of salt because I’m a professional corporate hater, though.