Welcome back! Today we have a quite heterogeneous lineup. From what could be Robotics’ ‘ChatGPT moment’ to the tragedy we are witnessing on bond markets, which could have a tremendous impact on AI’s ability to finance itself. We also include Cerebras’ new product, a DumpsterCluster, and more interesting stuff.

Enjoy!

As published by Generalist on August 19, the robotics startup has introduced GEN-1.5, a multimodal robot foundation model that can learn simple physical tasks from a single 3–12 second demonstration, without fine-tuning or gradient updates.

Generalist calls this approach “physical prompting”: it places a demonstration in the model’s context window, and the robot immediately attempts the inferred task with no training, succeeding in some very impressive cases.

Across 10 manipulation tasks (including opening jars, handling zippers and retrieving money from wallets), GEN-1.5 achieved an average 59% success rate from one demonstration. With around five minutes of task-specific data and 10 gradient updates, success increased to 83%. The company emphasizes that the tasks remain simple and short-horizon and that one-shot performance is still relatively brittle.

The model can also combine multiple demonstrated behaviors, imitate some demonstrations performed directly by humans, transfer demonstrations from simulation to real robots, and improvise with previously unseen tools. Generalist says it didn't explicitly train for these capabilities, but they emerged as it scaled pretraining on physical interaction data. GEN-1.5 has been training continuously for more than eight months.

This feels a lot like how Large Language Models (LLMs) were in 2020. Why?

The key is this physical prompting idea, where the model learns to do a task it hasn't seen before simply by showing a demonstration in the prompt.

This may sound familiar to those who have been in the space for longer, because that’s exactly what we achieved with LLMs back in 2020 in the seminal paper ‘Language Models are zero-shot learners’, published by OpenAI back when it had both OpenAI and Anthropic’s founders on the same team.

In that paper, they showed that, by scaling model sizes, you really didn’t have to retrain your model for every new task; just show the LLM a few examples, and the model would use those to learn how to do the task. Somehow, the model became capable of doing stuff it had never seen until that moment.

Now, robots are starting to do the same.

Six years later, LLMs are incredibly powerful AIs, and we’re now starting to see the same emerging behaviors in robotics.

The remaining question is reliability, which just isn’t there, as most tasks barely reach even 90% accuracy. As I’ve said multiple times, this isn't acceptable; you should aim for 99.9% or even 99.99% (three and four nines of reliability).

However, there’s hope: this is the same team that showed 99%+ success rates after task-specific post-training, which proves my earlier point: we’ll need task-specific training to reach the necessary production-grade reliability.

In layman’s terms, this is how I see things moving on:

While NVIDIA wants you buying new GPUs every year, researchers from Oxford and Cambridge built a 128-GPU cluster using nine-year-old V100s for just $32,000, called the DumpsterCluster, and found that, for the right workloads, it can actually deliver better performance per dollar than modern B200 systems.

How is that even possible? Simple: old GPUs have become ridiculously cheap, while their usefulness hasn’t fallen nearly as fast.

The V100 is obviously terrible compared to a B200 on raw specs, but AI is insanely capital-intensive, so what ultimately matters isn’t just how many tokens you generate, but how many tokens you generate per dollar invested.

I went much deeper into the economics, the engineering tricks behind the DumpsterCluster, and what this could mean for NVIDIA and the trillion-dollar AI CapEx boom in my full Medium article, which I posted yesterday and you can read for free here.

One thing I didn't cover in the Medium article is the key parallelization method they use, which explains how they offer better performance per dollar.

The secret is to split the model across GPUs by layer, a process called pipeline parallelism. Models are divided into neural network layers, so this is a pretty easy thing to do.

The key here is that if you keep each layer whole, the amount of data layers have to communicate is very small. This is particularly relevant here because the DumpsterCluster’s old GPUs have very poor communication bandwidth; the moment one GPU has to communicate with a GPU farther away, that bandwidth drops to 12.5 GB/s.

Therefore, if for whatever reason a GPU has to send another GPU, say, 25 GBs, it will take 2 seconds, a prohibitively long time considering all the work that has to be done to produce a single token. If a model has 40 layers, that is 80 seconds of delay for every single prediction, so each word would take a minute and a half to pop up on your screen.

But if you use pipeline parallelism—by the way, this is the same distribution mechanism Cerebras uses to break models in their clusters, for the same reason the researchers are doing it here—we dramatically reduce data communication requirements.

If we think about the Llama 3 70B model, this model has a hidden size of 8,192 numbers, the size of the internal vectors. If the sequence we’re processing has 100,000 tokens (~75,000 words), that is 100,000 × 8,192 = 819 million numbers we have to send between GPUs. Since that model uses BF16 precision (2 bytes per parameter), that's about 1.6 billion bytes of data, or 1.6 GB.

Luckily, if we use only pipeline parallelism, that’s all the data that has to be sent between GPUs during inference. As each GPU can communicate at 12.5 GB/s, that data will be sent between GPUs in (1.6/12.5) = 0.128 seconds, a communication penalty that can be easily hidden under computation (i.e., the computation will take longer than it takes to move that data in the first place, so that latency is essentially “hidden”).

See what they did here? They designed the workload to guarantee the bottleneck in their setting, GPU-to-GPU bandwidth, goes unnoticed. Pretty smart!

This has a pretty fascinating implication: GPU obsolescence might be much slower than NVIDIA would like you to believe. Old hardware isn’t as good as new hardware; it’s just becoming cheaper much faster than it becomes useless.

And if workloads can increasingly move between generations based on price/performance, AI compute starts looking more like a commodity. To me, this is actually good news. With the right workload, we can give hardware a longer life than it otherwise would.

On the other hand, this isn’t great news for hardware companies, which could see less demand for new chips as customers, tired of spending billions of dollars, decide to get “crafty.” Nonetheless, hyperscalers continue to use their V100 clusters.

However, as I explain in my Medium article, I highly doubt that will happen, because as token prices fall, volume—how many tokens they serve—is the only way they can guarantee solid revenue growth. Therefore, the ability to 10x throughput from one generation to the next, as Rubin showed with Blackwell, is too tempting to ignore.

This week we’ve experienced one of the biggest bond sell-offs lately, especially sovereign bonds, debt governments issue to finance their spending (most modern economies run fiscal deficits, meaning they spend more than they earn).

This has been pretty catastrophic and coincides with another debatable record set by the US, reaching $40 trillion (yes, with a T) in debt, reaching 123% of GDP, a very alarming value.

The situation is so concerning that the US Treasury has intervened and announced it will increase each purchase of US bonds from $2 billion to $4 billion. In layman’s terms, the US Treasury is going to start purchasing bonds to drop yields.

But unless you’re familiar with bond markets, you’re probably struggling to see the relationship with AI and why this matters so much to our beloved industry.

Bond prices and bond yields (the return of the bond) are inversely proportional; if bond prices rise, yields fall, and vice versa. Most investors simply memorize this, but it’s worth understanding why.

A government bond is just a form of debt governments use to borrow money. If I buy $1,000 of a 10-year bond with a 5% coupon, I’ll get paid $50 every year, and at maturity (year ten), I’ll also get the principal back. Thus, if I hold to maturity, I get my $1,000 back plus $50×10 = $500 in interest payments.

However, nothing requires me to hold the bond to maturity, and secondary markets allow these bonds to be traded at prices the buyer and seller agree on. For example, say I’ve purchased a bond but discover a new opportunity that might yield better returns, and I want to sell it.

The seller offers me $950 for the bond originally priced at $1,000, and I accept. I get back $950 (plus any interest payments I may have received while holding it). The seller buys a bond originally yielding 5% that now yields 5.26%, as the government is still paying $50 per year, but I purchased it for $950, $50 less than the original price.

That is why if bond prices fall (which is exactly what is happening now), yields go up. The US Government got freaked out (and for good reason) as yields for the 30-year bonds rose to the highest level since 2007, to 5.3%. If yields keep rising, it will be more expensive for the US Government to borrow new money, so it urgently needs to bring them down.

Therefore, the immediate plan is to repurchase a lot of that long-term debt, reducing supply, causing bond prices to rise, and thus yields to fall.

In Spain, we call this “pan para hoy, hambre para mañana” or “bread for today, hunger for tomorrow,” which is basically what we’re doing here: we’re reducing debt costs for today but doing so by increasing our debt burden for the future (as the US Treasury has to emit new short-term bonds to get liquidity to make these purchases).

But is the situation really that bad? Well, yes.

What’s perhaps more striking is that debt servicing (paying back the debt) already represents a top-three spending category. Yes, the USG spends on paying back debt among its top spending categories, which naturally means the only way it can pay that debt is… you guessed it, more debt!

At this rate, debt servicing could be the US’s largest spending category ahead of Social Security by 2028.

The truth is, for one too many years, alternating US Governments (and any government from a modern economy except Germany in the last 30 years) have failed to reduce the debt burden, because the only way to do that is the ugly way: cutting.

And cutting takes away your power.

This is the exact same issue all major governments (e.g, UK, France, Spain, Canada, Italy, and more) are facing: are we going to be the ugly ducklings doing the painful cutting, or are we going to make the double the problem and give it to the poor bastard that’s coming next?

And what does this have to do with AI? Well, two things, one for the short term and another for the long term.

Starting with the latter, AI is one of the last hopes the US has to avoid a debt spiral. Debt won’t really run out of control as long as it remains a “reasonable” percentage of GDP (I wouldn’t consider being larger than GDP good, but it’s just the reality for most developed economies these days), so the hope is that AI translates into an absurd amount of production growth that “pays” for the debt.

I’ve already made my case for why this doesn’t seem likely based on history, but I think we can all agree it would be great if I'm wrong and AI really transforms GDP soon.

The other relationship's effects will be felt much sooner, though. In short, AI is actually making the problem worse.

But how? One problem with the USG’s debt burden is that its yields are rising simply because of opportunity cost.

Why buy US Government bonds and get “stuck” with them for who knows how long at 5% when I can purchase Alphabet (Google) corporate bonds at almost 7%, paying me more while holding debt from an entity with pristine credit quality?

The truth is, this is a self-reinforcing mechanism for both: rising yields on government bonds naturally push Big Tech yields upward (as they are riskier borrowers for the most part), which… causes government yields to rise more, and the cycle self-reinforces.

Nonetheless, companies like Bank of America say the surge has added about 0.3 percentage points to the government’s 10-year yield, and the diverging trends are directly visible:

The problem? This is only going to get worse, as AI will need much more debt over the next few years, potentially trillions of dollars, and god knows what prices these companies will be paying a year from now.

Meanwhile, companies like CoreWeave have been in junk-bond territory for quite some time, and others like Oracle are BBB-, the worst investment-grade rating, literally bordering junk-level territory.

Time for the Hyperscalers to join this group of dubious reputation? Won’t happen anytime soon… I hope.

As reported by WallstreetCN on August 20, Alibaba said on its earnings call that AI-related products have reached RMB49.5 billion ($7.3 billion) in annualized revenue, representing 35% of Alibaba Cloud’s external commercial revenue. Cloud revenue grew 45% year over year, while AI-related product revenue posted triple-digit growth for the 12th consecutive quarter.

Moreover, CEO Eddie Wu said Alibaba’s in-house T-Head Zhenwu M890 AI chip has entered large-scale commercialization, with cloud instances based on the chip now on sale and production expected to ramp further in the second half of 2026. Alibaba said its AI-related products also carry higher gross margins than its cloud average.

Alibaba has invested about RMB190 billion ($28 billion) so far under the RMB380 billion three-year AI infrastructure program announced in 2025. Crucially, CFO Toby Xu said current AI compute investments are expected to pay back in roughly three years, while Wu said improving margins could eventually shorten that period to around two to two-and-a-half years. The spending is already weighing on results: quarterly capex rose 75% to RMB67.7 billion ($10 billion), while net profit fell sharply despite 9% revenue growth.

We have yet another strong piece of evidence that AI is profitable, although it has not yet proven cash-flow generating potential. Or has it?

If Alibaba’s claims are true, it’s not only profitable, but it might be very profitable and generate meaningful cash flow returns, because three-year paybacks mean AI hardware generates returns far faster than it depreciates.

With the debt issue worsening (see news above) and several Big Tech companies falling into negative free cash flow, AI needs to start paying back. And fast. Enough of empty promises and delusions of grandeur; AI has to start proving its value and generating real cash.

Otherwise, it’s unclear how long markets will remain willing to pay for the lights to stay on, and debt markets are starting to show signs they might give up.

Over the last few days, there's been a lot of noise about falling token prices as a sign of weak demand. If people want less AI, you drop prices to revitalize demand, right?

Well, could be, except for the fact that data easily contradicts this, and what we’re seeing is simply the Jevons Paradox all over again. Algorithmic progress drives lower token prices through competition, and then demand increases because of the efficiency gains.

We know this because as token prices continued to fall, GPU rentals started to rise, meaning lowering prices is actually leading to more customer demand.

Demand is alive and well. However…

The remaining unanswered question: to what extent is demand investor-subsidized?

People see this and immediately assume all this demand is healthy. That is, real customers, using AI. However, this ignores the elephant in the room: much of this demand comes from AI companies that have raised money from investors and use it to train new models, run experiments, and the like.

While I obviously like to see greater demand at lower prices, which signals real demand, I believe the overall market still depends heavily on investors pouring money into the space.

In my view, demand is evolving radically and asymmetrically, as I explained in the last Leaders’ newsletter, as AI continues to disappoint in its ability to monetize the average individual, who is generally not interested, as shown by this Ramp graph.

Ramp also shows that some smaller cohorts are spending like crazy. So, here’s the thing: I’m becoming very convinced that the AI bills will be paid by a very small cohort of people, and the vast majority of individuals will not pay a single dime.

This is fine as long as the maths ‘maths out’, but places enormous pressure on AI companies to keep token prices high or manage to make these people spend an absurd amount of tokens.

But what if the top 1% starts budgeting? Jane Street might not care and pay 20x the average token price to have access to “instant” AI models running at thousands of tokens per second, but Uber might not be as willing to spend thousands per customer.

Those dynamics are all but clear and will materially decide whether AI booms or busts. But, to me, the lesson is strikingly clear:

In fact, I think most AI inference will be served for free (not only is ChatGPT free, but other companies like Replit are starting to offer free products), and these products and services will be embedded in “free” software like Google search, Meta’s ad targeting, or OpenAI’s ChatGPT freemium product, where you are the product because you’re going to be fed ads in exchange for not paying a dime.

The world has yet to see a billion-customer-scale software product that wasn’t free (i.e., running ads on free customers). And nothing tells me AI will be different.

As published by OpenAI, the company is previewing Private Safety Processing, a system designed to let eligible API customers keep Zero Data Retention (ZDR) while enabling stronger safety monitoring across multiple related interactions.

Under ZDR, OpenAI says it does not retain prompts and model responses after processing, does not make them available to its personnel, and does not use enterprise data for training unless customers opt in.

The new system is intended to detect misuse patterns that may only emerge over several interactions. Customer content can remain on infrastructure controlled by the customer, or eventually be stored by OpenAI while encrypted with customer-controlled keys. Automated systems would send OpenAI only limited safety signals, without exposing the underlying prompts or responses to staff.

OpenAI says it is testing Private Safety Processing with early customers and plans to begin rolling it out and publish a technical white paper in September 2026. One exception to ZDR remains legally required handling of images flagged as potential child sexual abuse material, which may be retained for manual review and reporting.

Can’t wait to see European bureaucrats objecting to this because they want to see the data. Jokes aside, frontier labs now have to deal with a lot of enterprises asking questions about how they handle user data.

Of course, labs would love to see your data, but this has become a deal breaker for basically anyone, especially considering that AI incumbents themselves, companies like Microsoft or Palantir, are telling customers “yeah, you should not sign that deal with OpenAI.”

Personally, I think sovereign AI enterprises owning the entire stack, including the model (i.e., using mostly open models), is inevitable. This news might delay the transition, but I really don’t see frontier models in organizations beyond marginal, quality-maximizing work.

Don’t get me wrong, that could still be a huge business for them (e.g., companies will pay top dollar to secure their network using AI’s alleged cyber capabilities) but if you’re still using Claude/ChatGPT throughout your org without a clear plan to transition to open models at least partially, you might run into serious cost issues if labs suddenly eliminate subscriptions (I don’t think they will in the short term, but I wouldn’t be comfortable without an ongoing transition plan because this could change anyday soon), especially once Anthropic IPOs and gets the liquidity it needs from retail investors.

After that event, they are no longer having to pretend they care about you and will instead be under a lot of pressure to raise margins, something diagonally opposed to what subscriptions produce. I continue to doubt subscriptions will have a long life in the enterprise.

Cerebras has released its new hardware platform, the CS-4, and it shows incredible performance, promising 2,000 tokens/second/user for frontier models and in the multiple thousands for smaller models.

The CS-4 basically doubles the specs of the previous platform, the CS-3. This means they are essentially promising double the performance. If your model was running at 500 tokens/second/user in a CS-3, it will now run at 1,000 tokens/second/user.

Incredibly, this is despite the wafers (i.e., the chips) being pretty much identical between generations; the chip has not changed.

Is this magic? No, it’s much easier and boring; they’ve doubled the power to double the clock frequency. In other words, every chip now does twice the “work” per second, doubling everything else.

This means that each wafer now requires ~46 kW of power (previous wafers required 23kW). If we have three of them per CS-4, that is roughly 140kW per CS-4, in a similar range to NVIDIA’s top servers today.

So we’re getting many more tokens for roughly the same power? By the looks of it, this should be NVIDIA’s death as it beats the latter’s performance by an order of magnitude.

However, this is nonsense, and NVIDIA will be just fine, because this is an apples-to-oranges comparison. Why? Because both platforms serve very different purposes.

I won’t elaborate too much for the sake of length, but Cerebras achieves this throughput on a very particular instance: inference decoding with just a few users.

In layman’s terms, Cerebras is designed to serve AI models to a small group of people really fast. If you try to use Cerebras chips to serve hundreds of users, performance will be absolutely terrible.

What this means is that Cerebras’ platform can never, and I mean never, run in isolation; it has to be paired with GPUs so that each accelerator focuses on the part they shine in, the already pretty well-known idea of disaggregated inference.

Moreover, Cerebras hardware is extremely pricey relative to the revenues you can get from them (as mentioned, if you try to serve a lot of users with Cerebras chips you’re going to offer a really bad experience, which dramatically reduces your addressable revenue), so the only way to make the business case pan out is to charge a lot.

That is, Cerebras is a premium inference play: serve a very small number of tokens overall, but do so very fast for the few customers you can serve, and charge them a fortune.

Cerebras’ entire existence depends on the premium inference market being large enough.

If so, Cerebras becomes a very appetizing option. Luckily for them, based on what we’re seeing from the Ramp data I referenced in the markets section, that’s exactly how the AI demand distribution is looking; a small group of people willing to pay a lot of money.

If that view materializes, which I think it will, Cerebras is right on the money—I don’t own any Cerebras shares at the time of writing (and never have).

As published by Anthropic, Claude Mythos 5 is being expanded to more cybersecurity defenders while Anthropic keeps tighter controls on direct model access.

Mythos 5 is now powering Claude Security scans for Enterprise customers and will also be integrated into partner security products, where users receive limited outputs such as vulnerability alerts or suggested patches rather than unrestricted access to the model.

Anthropic says Claude Security can scan repositories, classify vulnerabilities, assign confidence and severity ratings, and suggest fixes, with all patches requiring human approval. The company is also launching the Defender Advantage Fund, providing $35 million in credits to organizations working on open-source vulnerability detection and remediation.

Anthropic also plans to expand its Cyber Verification Program, giving approved security teams broader access to Mythos-class defensive capabilities and fewer restrictions when using Claude Opus and Sonnet for legitimate cybersecurity work.

This is interesting because it’s the first time average enterprises are getting access to Mythos 5, not Fable (the heavily guardrailed version), so it should be pretty darn powerful.

It seems like finally all those concerns about releasing top models to average users is suddenly no longer an issue casually when their competition is more fierce, with OpenAI allegedly growing at a faster rate than them lately, with GPT-5.6 Sol being considered by many the state-of-the-art while they are decreasing its price by 20% for three months and with companies like SpaceXAI or Meta offering models that blow Anthropic’s models out of the water in terms of intelligence per cost.

Anthropic is obviously realizing it can’t sustain those prices unless it releases the best it can offer, showing that, just like any other company, when push comes to shove, money beats ethics.