Welcome back! This week, we have lots to talk about. From AI satellites and an AI company sabotaging customers to what I believe is the future of edge AI: diffusion models.

Enjoy!

Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9, introducing a new Mythos-class model family for advanced coding, knowledge work, vision, scientific research, and long-context tasks.

The innovative aspect is that Fable 5 includes additional safeguards relative to Mythos in areas such as cybersecurity, biology, chemistry, and model distillation, due to risks perceived by Anthropic.

These safeguards are applied using classifiers that detect certain requests in those categories. When detected, the response will instead be handled by Claude Opus 4.8, and users will be informed when this occurs.

Only the “selected few” with access to Mythos, unlike the peasants you and I are, will be able to use frontier-level capabilities to make biology or cybersecurity work.

Much worse, for the particular topic of Frontier LLM development, Anthropic has done something unprecedented: they purposely downgrade, or dumbify the model so that it provides worse answers… without telling you.

Have you ever met a company willing to sabotage your responses, without telling you, just because they disliked what you asked?

This announcement caused massive controversy, and today they actually dialed it back: they won’t be dumbifying Fable anymore and will instead explicitly downgrade you to Opus 4.8, like in the other topics… or so they say (I don’t trust their word at this point).

Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens.

I don’t want to say I told… but I did. If you’re a regular of this newsletter, you know I’ve not been shy about calling Anthropic out for perhaps years.

I never trusted them, and the world finally understands why.

It all started once they began with their fearmongering about AI destroying humans, equating them to nuclear bombs, and all that condescending “I’m creating a God, only we should do it” stuff without giving a shred of proof.

I will say it stuck with most people for a while. But I didn’t buy it because I understand the technology and its limitations, and I knew from the start it was all a marketing-slash-regulatory-capture strategy.

To me, the most unforgivable thing is not that they decide I’m not worthy of using the frontier models, but the fact that a company trained on all of our data, while training models based on architectures developed by Google, not by them, only to pull up the ladder behind them once they were ahead, is just upsetting.

Just picture a world where these guys get to decide who can use, and when, the most powerful technology humans have probably ever built. I don’t want to live in that world in the same way I wouldn’t want to live in Turkmenistan’s dictatorship.

But I’m a positive guy, so I have positive takeaways:

SpaceX, the new 2-trillion company (see below) has presented what would be its first AI satellite running computing from space. It’s a massive satellite, more than 70 yards long, with computing power of up to 150kW at peak and capable of sustaining 120kW on average.

For reference, that is roughly the electrical draw of a dense Blackwell Ultra-class AI rack on Earth.

SpaceX is quite literally the definition of a moonshot. I haven’t bought into the IPO, but if they get this right, SpaceX is the ultimate AI play: they design the chips, they will manufacture them too, they have the rockets to push them into space, and they train the AIs that go into them while also being capable of supplying compute to other players like Anthropic, Cursor (which they acquired), or Google.

It’s a great company, but in my view outrageously overvalued (more on this below).

Continuing with the star of the week, as published by Reuters, SpaceX has signed a multi-year cloud services agreement with Google, under which Google will pay SpaceX $920 million per month from October 2026 through June 2029 for access to AI computing capacity.

The deal gives Google access to about 110,000 Nvidia GPUs, plus CPUs, memory, and related infrastructure.

With this deal, SpaceX’s RPO compute backlog (the “agreed” revenue commitments it has signed with customers) amounts to $70 billion over roughly the next three years between the Anthropic and Google deals.

For reference, SpaceX’s core business until recently was Starlink (Internet satellites), which has an ARR of $13.6 billion based on Q1 revenues. This means the AI business, in just two deals, is already twice as large.

The caveat, naturally, is margins.

While the Starlink business has a 38% operating margin (meaning it still makes money after subtracting production and operating costs), the AI segment, despite its amazing revenue growth, is enormously unprofitable once we account for the huge depreciation costs associated with AI hardware and research & development costs (researcher salaries and training costs).

Recursive AI, a neo AI startup, has released early results from an automated AI research system that can propose ideas, implement them, run experiments, validate results, and use prior results to select subsequent experiments.

Recursive says the system reached state-of-the-art results on three benchmarks: fixed-budget small language model training, small-model training speed, and GPU kernel optimization.

The system tested changes in model architecture, optimizer behavior, embeddings, attention precision, compiler settings, and fused GPU kernels. Recursive says it also screened results for reward hacks and variance before treating them as improvements. They are also open-sourcing artifacts from the runs so others can inspect and build on the outputs.

This is the way. Research that is created in the open to benefit all, not whatever Anthropic’s leadership with a God complex intends to do.

These RSI first steps are just that, minor wins that could one day lead us to a future where AIs indeed self-improve.

The fact that AIs can modify themselves can be viewed with fear, sure, but that’s precisely why this needs to be done in the open, not behind walls of regulation and money, as some believe it must be.

Money and power corrupt, so it’s vital that a technology that could one day become unstoppable be built in the open.

SpaceX went public today, pricing its IPO at $135 per share and raising roughly $75 billion at an initial valuation of about $1.77 trillion.

Shares were indicated to open higher, around $171, implying a valuation above $2 trillion. The stock is trading at $2.1 trillion, the same market value as TSMC, at the time of writing.

For reference, this makes SpaceX the seventh-largest company in the world by market cap, just below TSMC and ahead of companies like Meta or Broadcom. Its value is greater than that of JPMorgan and Walmart combined.

Absolute madness. SpaceX trading higher than companies with:

It will be worth more than a potential corporation with almost 50 times the revenue (Walmart + JPMorgan).

And look, I see the appeal in the company’s business. They have quite literally everything: a verticalized AI play, a monopoly on rockets and satellites… except the most important thing: profits.

The deals with Anthropic and Google really help (the latter explained below) really help paint an improved picture, adding $70+ billion to future revenues, but will do so across three years from now, and will depend, especially in the case of Anthorpic, of cash flows that Anthropic’s business will not be able to guarantee (meaning, that money will come from investors or debt).

I’ve maintained for a while now that this year’s IPOs, especially the upcoming Anthropic and OpenAI IPOs, will determine what we make of AI’s future.

But seeing the—quite frankly delusional—fervency around SpaceX’s IPO makes me think that AI is going to be just fine because retail investors are willing to throw their money down the drain.

According to the Wall Street Journal, OpenAI is considering cutting token prices to put pressure on Anthropic, which would open a new chapter in the price wars that have been on a truce for more than a year now (token prices have barely moved in recent times).

This is not something OpenAI can afford to do, but something to pressure Anthropic, which has a much smaller customer base, to reduce prices at a time when OpenAI’s Codex is stealing a lot of users from Anthropic due to the higher rate limits (because OpenAI has long had a much greater amount of compute than Anthropic, somethng the latter has sort of fixed in the last two months, but paying a high price for it).

While I understand OpenAI’s reasons for this, these companies are really never going to make money, do they?

The reason is quite simple: it’s a commoditized technology (otherwise, price wars wouldn’t be a thing) running on non-commoditized hardware (the hardware AI needs has some of the highest gross margins on the planet).

Particularly daunting is the topic of subscriptions. SemiAnalysis ran a series of tests yesterday and found that subscriptions are, quite literally, cash-burning machines for these Labs, with spend reaching $14k for a $200 subscription.

This shouldn’t be new to you, considering I’ve long talked about the unprofitability of subscriptions.

But why? At the end of the day, the rationale is pretty straightforward:

AI is a technology with high marginal acquisition costs, meaning each individual user can generate outsized costs due to the high hardware intensity of the average workload (as we discussed in our last issue, the real problem is capital costs, not really inference costs).

This means OpenAI or Anthropic can predict the profitability of a given subscription; you could make 95% gross margins, or -500%, depending on the user. This breaks the historical “software contract,” in which, under CPU-based regimes, marginal costs were negligible.

Pre-AI, for SAP, selling a $30/month seat would require, at most, $3-4 in production costs, give or take, guaranteeing +85% gross margins no matter how “eager” that user was.

As long as that remains true (hint: it will), subscriptions don’t make any sense, and usage-based pricing is the only path to profitability.

This is why I enter every single board room I go into with this tattooed on my head: assume subscriptions will disappear; build an IT government model that assumes usage-based pricing on a cost-plus basis, where every token counts.

OpenAI might make the subsidized era last a little longer, but it doesn’t change the future, which is a pay-as-you-go model.

As for consumers, where usage-based models rarely work, it’s a non-issue because most consumer tasks (searching for stuff, asking questions, editing videos, etc.) will mostly run on edge hardware in one/two years’ time (more on this later in this newsletter).

Ramp has launched an AI-applied solutions service, in which Ramp engineers embed with enterprise finance teams to build customized AI agents and workflows on Ramp’s platform.

Ramp says the offering is aimed at companies that are increasing AI spending but struggling to show measurable returns.

It cites internal customer data showing that AI token spend across its 70,000+ customers has risen 13x since January 2025, while only 21% report measurable results.

Ramp says the service is model-agnostic, routing workflows to different AI models based on performance, cost, and task fit, rather than locking customers into a single provider.

The company says deployments are designed to ship into production “in weeks,” with customers able to take ownership after handoff or keep Ramp involved as the system expands.

Even though I’m not a Ramp customer (never tried the tool), this is the exact playbook a software company will have to follow to survive.

I’ve maintained for a long time two things about software’s future:

Ramp ticks both boxes with their decision. On the one hand, it’s clearly becoming an AI-native company, actively pushing research and fine-tuned models.

But instead of just waiting for customers to build their own bespoke solutions to replace the Ramp license, they offer that customization on top of the Ramp platform, providing the “be spokedness” customers want while still creating client lock-in on their product.

If you’re an investor looking to see what software companies will exist in five years, this is the type of sign you should be looking for: companies like Palantir or Ramp that, instead of offering a generalist software, proactively move into the customer and become their bespoke solution.

Google has released a new Gemma 4 variant that can run on consumer hardware, delivering strong performance for its size.

But the fascinating thing is the architecture itself: it’s a diffusion model, hence the name DiffusionGemma.

In other words, unlike standard Large Language Models (LLMs), which generate one token (e.g., a word) every round, generating an ‘autoregressive’ sequence of tokens, Diffusion models take a more immediate approach.

They depart from a noisy overall picture, and iteratively “denoise” the slate to “unearth” the response.

I’ve always thought of this intuitively as sculpting: you take a huge block of marble and “unearth” the “hidden” statue by chiseling away the excess. As the great artist Michelangelo once said:

This means diffusion models progressively transform what’s essentially noise into an actual result by performing several ‘denoising’ updates.

However, diffusion models introduce an unequivocal trade-off: they are much faster because they generate results globally rather than sequentially, but in most cases this implies a decrease in performance.

Google itself mentions this: “For applications demanding maximum quality, we recommend deploying standard Gemma4”.

But DiffusionGemma and other diffusion models do introduce one vital aspect that makes me particularly bullish about them: they massively reduce the memory bottleneck.

Too long to explain here (in case you want the longer explanation, read here), but I’ll try my best to simplify. Computers work by moving data into a processor, which performs a series of operations, and most of the results and the input data are sent back to memory, creating a back and forth between the memory and the compute.

This means that both are needed, and the slowest of the two is the bottleneck. In AI, especially inference, memory is the bottleneck, which is why AI is famously defined as “memory-bound.” In practice, this means that, on average, processors are somewhat idle, or “waiting” for data to arrive.

For companies that make money by producing tokens, that idle time translates literally into lost revenue.

The reason inference is so bound by memory is easy to see. Think about how ChatGPT works, generating one word at a time. In practice, this means you push the model into the processor, the processor decides the next word, and the process repeats.

The issue is that models are so large they break down before being pushed, so to make way for the next part of the model, we need to extract the previous part, increasing the amount of data moved in and out.

Naturally, this means that if data movement is the bottleneck, the way to squeeze the most performance is to make the most of every byte of data we push into the chip.

That is, if a GPU can allegedly perform 100 operations per byte of data the processor sees, the goal is to ensure we stay as close to that value as possible. This way, we guarantee that processors are running at a good pace.

And what does all this have to do with diffusion? Simple: while we have to do this entire data dance to generate a single token in an autoregressive model like ChatGPT, a diffusion model updates 256 tokens at each step.

Careful, I’m not saying that every prediction pass churns 256 tokens, but a step in the denoising process. However, that results in extremely fast generation because once the denoising updates finish, you automatically get 256 tokens.

For example, for two identical models, one autoregressive and one diffusion, if the denoising steps are 30 (as in the Sudoku example above), while an autoregressive LLM generates 30 tokens, DiffusionGemma generates 256.

The reason I’m so bullish on diffusion models is that they will play a crucial role in edge hardware. Our smartphones and laptops don’t have nearly as much capacity for fast token generation as cloud servers do, and thus require algorithmic improvements to churn tokens faster.

It’s not the intelligence level that makes edge models hard to adopt; it’s how slow they are. Diffusion models can massively increase token generation speed, making adoption much easier.

As they reach “good enough” thresholds, I believe they are going to be representative of a huge portion of the world’s generated tokens.

One of the most powerful yet often-overlooked abilities of frontier AIs is reverse engineering: they can examine a product or service and determine how it works. For example, they can look at screenshots from an app and generate the source code without having seen it before.

An X user has connected their Whoop app, which controls a wristband that tracks heart rate and other metrics to assess stress or sleep quality, to Calendar, so it can track which meetings (and, thus, people) give them the most stress.

To do this, they used the new Claude Fable to reverse-engineer the Whoop so that it could pull per-minute heart rate data and, in this way, associate what people give this man with more stress.

What willpower and boredom do to someone, right?

More than the funny use of the technology, my takeaway is how easy it’s going to be to build apps on the fly.

Not necessarily things you can sell (with more apps available, getting attention will be harder than ever), but creating software as a way to solve personal problems, not necessarily as a product or service for others.

New project? An app to track progress with the particularities of this one. New son's hobby? Build an app that helps your kid.

The opportunities will be endless, and so will the cybersecurity risks, but we’ll leave that for another day.

Harvey, the AI legal startup, announced a partnership with Trajectory Labs to post-train NVIDIA’s Nemotron 3 Ultra for legal-agent work, achieving frontier-level results in just 24 hours of fine-tuning on NVIDIA's open models using reinforcement learning (RL) and implying a 50x cost reduction.

According to Harvey, the base Nemotron 3 Ultra model scored 0% on LAB’s all-pass metric before post-training. After less than 24 hours of post-training, the model reached 5.8% all-pass. That sounds mediocre, but it’s almost as good as Claude Opus 4.6, a frontier-ish model.

I’ve been saying for years that fine-tuning will play a vital role for enterprises, taking an open, non-frontier model and turning it into a frontier model (for that particular task), training it on task data.

I’ve long maintained that once enterprises start training models on company data, OpenAI and Anthropic will have a hard time selling into businesses, because companies will have the option to use free or cheap models that can be run securely within their orgs and achieve frontier-level performance without the added premium costs.

As I always say, enterprises don’t need generalist savant models; they need models that do a given task well, even if training them to do that well makes them worse in other areas. Who cares? Just pick another model for the other tasks.

The frontier market will still exist, as some critical use cases require frontier-level performance. Think coding, drug discovery, science, cybersecurity (to some extent, not really true), or maths; in those areas, you want to maximize “intelligence.” But for everything else, what you want to maximize is “intelligence-per-dollar”.

And in that regard, Anthropic/OpenAI models are nowhere to be seen:

It was inevitable, but I can’t say I’m not spooked either way. Chinese company UbTech has presented its humanoid companion for pre-sale, getting 3,000 orders in a week. The robots are human-like and are meant to serve as companions.

How far has humanity fallen to requiring non-human humanoids to not feel lonely (or to get laid, which is even more concerning)?

I can understand having a humanoid handling packages in a factory; I don’t think that’s anyone’s dream job. But actively trying to assume the roles of other humans in basic human-to-human relationships is Kafkaesque.