Welcome back! Lots to talk about this week, from OpenAI’s latest controversy, an AI Scientist that can work for 12 hours straight and advanced research by six months (allegedly), a Chinese SOTA model that might have definitely closed the gap with the US (even more so than DeepSeek), Google’s Siri deal, the insane numbers Morgan Stanley believes Hyperscalers will spend, and more.

Enjoy!

On Tuesday, we took a look at a long and exciting list of news, ranging from new deals by OpenAI to how Microsoft revealed OpenAI’s losses (they are huge), why AI robots are still far from being remotely useful, a concerning trend amongst Hyperscalers’ huge AI bets, and more.

And we concluded with my newest stock purchase, a company I believe is outrageously undervalued and potentially a huge stock winner in 2026, with TSMC/ASML-level bottleneck in the entire AI supply chain, all 2026 revenue pre-sold… Yet deeply ignored by markets amidst what could soon be one of the most significant supply squeezes in history.

Transform your team’s static training materials into dynamic, engaging video guides with Guidde.

Here’s what you’ll love about Guidde:

1️⃣ Easy to Create: Turn PDFs or manuals into stunning video tutorials with a single click.
2️⃣ Easy to Update: Update video content in seconds to keep your training materials relevant.
3️⃣ Easy to Localize: Generate multilingual guides to ensure accessibility for global teams.

Empower your teammates with interactive learning.

And the best part? The browser extension is 100% free.

Check out Guidde

In an interview with the WSJ, Sarah Friar, OpenAI’s CFO, publicly asked for a federal backstop for OpenAI to guarantee a safer, lower-risk way of guaranteeing AI infrastructure and chip buildouts; a free bailout card basically.

In the short clip, she makes the case for it, saying that these buildouts have become too risky for banks (duh), especially considering the underlying asset, the accelerators, depreciate pretty fast (the official lifecycle for a GPU is three years, although prior generations like Ampere are still running 5 years since release), making lenders skeptical of committing too much money.

Therefore, she requests a federal backstop guarantee from the US Government, meaning the government would provide liquidity to banks in the event they need it, to ensure the buildout is executed.

Unbelievably wild take, but not surprising at all. This is what we’ve been saying for weeks now: OpenAI is making itself too large to fail.

Although she has backtracked on the claim (and so has Sam), it’s clear she meant it, which suggests OpenAI is running into newfound struggles to finance its overly ambitious $1.4 trillion buildout plan. Nothing we didn’t already know (NVIDIA’s $100 billion “loan” was all we needed to know), but I was surprised to see them taking this stance publicly so soon.

I’m speculating here, but these comments and Sam Altman’s recent stressed and defensive interview in the B2G podcast with Brad (OpenAI investor by the way) signal to me that things aren’t going perfectly at all.

And in case you’re wondering, this differs from the US government’s $10 billion stake in Intel. One thing is to provide liquidity to a company that’s clearly the US’s only way of getting a chip foundry champion (an obvious national security interest and even potentially generating revenues for US tax payers in the future); another is to tell the financial institutions in America not to say no to OpenAI because they’ll bail out the company to the tune of a potential $1.4 trillion debt.

It’s already challenging enough to justify the government investing in a private company, but another story is effectively privatizing the profits while socializing the losses.

It’s not up to the US taxpayer to pay for OpenAI’s struggles, so let’s hope the Trump Administration does not fall for that, especially when the taxpayer won’t be receiving profits if OpenAI succeeds.

EpochAI has released, free of charge, an interesting tool that enables you to track the various giga-scale data center buildouts being constructed in the US.

You can view their locations, as well as satellite data of the site itself, and visualize the different buildings, cooling systems, transformer substations, and more.

The tool also includes sources for power, which is quite handy in case you need to verify. Ultimately, these figures are based on estimates, rather than official data.

As you can see below, several data centers of giga-scale size are already scheduled to go live early next year. And by early 2028, we could see $100 billion data centers with more than 3.3 GW of power capacity deployed, like Microsoft’s Chicago Fairwater site.

Even though this is not based on official data, just by looking at the magnitude of the site with the satellite data is enough to know how much power they’ll be working on. For example, you can count the number of cooling chillers they have and derive the power.

For instance, Colossus 2 has 129 chillers (shown below in the top side of the image). The combined cooling capacity is 200 MW.

Assuming industry rule of thumb for air-cooling, where ~40% of the total power the server draws is required for it (liquid cooling requires much less, Supermicro quotes up to 40% reduction), and knowing these sites have NVIDIA’s GB200 NVL72 servers, each requiring, on average ~130 kW (this numbers on the OEM deploying them), that means each server requires ~52kW of cooling.

If we have 200MW of it, that means we can deploy ~3,800 of these servers, which at 130KW, gives a rough value of 0.5 GW for that particular building's AI compute power, which makes sense considering Colossus 2 is a giga-scale data center with more buildings on site.

All things considered, a great resource to track data center buildouts and a more visual representation of the 13 projects that are keeping the US’s GDP growing.

According to recent sources, Google might be nearing another investment round into Anthropic, this time valuing the company at a staggering $350 billion. Google is already the second majority owner of Anthropic behind Amazon.

The reason is none other than securing compute for their upcoming official launch as a cloud chip provider, as they near the announcement that they will start renting TPUs to customers.

TPUs are Google’s own AI accelerator. But since its first generation, Google has always used TPUs internally (to train and run models like Gemini), rather than offering that compute via a cloud offering. Now, it seems they believe they can compete with NVIDIA, which would open a new revenue segment for the already pretty diversified (revenue-wise) giant.

Thus, this investment in Anthropic gives them a way to formalize and secure a client, even if, once again, it’s a self-generated revenue play, meaning it’s not real revenue but Google spending money on its own TPUs through Anthropic. Either way, between this and the Siri announcement, things are looking pretty incredible for Google.

According to a leaked Morgan Stanley analysis (I couldn’t get a hold of the report, sorry), the top six Hyperscalers (including CoreWeave) are expected to deploy almost $600 billion in 2026 and up to $700 billion combined in 2027, most of which, of course, will go to AI, truly insane numbers.

As always, the question that follows is: Will revenues follow? Will these companies make a return on this investment? These are the obvious questions… but what if they are not the correct ones?

May I present to you my new thesis on what is going on: it doesn’t matter.

Which is to say, they aren’t building this to make it profitable (at least it is not a non-negotiable result); they are building this to make themselves indispensable.

If AI becomes the foundation of all meaningful economic activity, whether it generates a return or not is irrelevant if it becomes the infrastructure of that technology.

This may seem stupid when you consider how investors value companies (primarily based on earnings). However, there are several examples of companies with slim-thin profits, or even unprofitable for years, that are still extremely valuable because the world would “stop” without them.

Companies like Walmart and Amazon barely make money (Amazon, I mean the non-AWS business, which is still at the heart of what makes Amazon truly valuable as a company), but no investor in their right mind would not be invested in companies that are indispensable to society; that’s what makes them valuable, regardless of whether they are making cash or not.

This is the only reason I can see why these companies would be willing to take such huge risks. Most of them have extremely profitable cash cows and could use the money to invest in future high-margin businesses. Instead, they are going into a high-risk/”low-reward” ugly business of building data centers.

Of course, enthusiasts will tell you, “No, bro, they are investing to build AGI.” But I mean, come on, shut up. No serious Big Tech CEO is truly signing off on these huge commitments on a loose promise to get AGI. However, what they all seem to agree on is that the technology will play a crucial role in society, whether or not it involves AGI, whatever that term means at this point.

But another reasonable pushback on my thesis is my claim that it’s low margin. How are data centers considered to have low margins? And the answer is that your viewing this from the perspective of CPU-based clouds. Of course, CPU-based data centers are high-margin (just consider their cloud divisions).

We talked about this a while back in our newsletter regarding who would make money and who wouldn’t in AI, but the point here is that AI clouds are low margin because in order to provide your service, you need to generate tokens, which means greater energy costs that grow proportionally to greater revenues.

Let me put this more clearly: non-AI clouds have extremely low marginal costs (i.e., at a sufficient scale, adding a new customer is essentially free), while AI clouds have sizeable marginal costs; adding a new customer might require ramping up an entire GPU server for them at peak loads.

So, not only are they investing in a business with an unclear path to revenues (let alone profits), but it’s one that forces you into a never-ending spree of new buildouts (there’s a reason Sam Altman wants to create a “GW factory”; they aim to deploy a new GW of AI compute every week at one point). Make it make sense.

At this point, the only reason I see for doing this is that they view it as a way to establish the foundation on which all future economic value is built. If that happens, who cares if they barely make money? The world will come to a standstill if these guys stop producing AI outputs. If that vision becomes true, profits will no longer be relevant in that dichotomy.

A Google DeepMind team, led by Pushmeet Kohli, collaborated with mathematicians Terence Tao and Javier Gómez-Serrano to leverage AI agents (AlphaEvolve, AlphaProof, and Gemini Deep Think) for advancing math research. And it worked.

They used AlphaEvolve, an AI system from Google DeepMind, which I covered back in May, designed to automate scientific and mathematical discovery. It utilizes large language models, such as Gemini, as part of an iterative search loop that writes, tests, and improves code to solve open problems.

Rather than directly generating new solutions, AlphaEvolve evolves the search process itself—optimizing how code explores possible solutions and gradually improving its heuristics based on performance feedback.

In layman’s terms, it’s not actively proposing the exact solution, but running search algorithms at scale that discover it. Conceptually, AlphaEvolve demonstrates that AI doesn’t need human-like understanding or reasoning to make discoveries. Instead, by scaling recursive, compute-heavy search guided by LLMs, it can uncover results beyond human reach.

AlphaEvolve discovered new results across various problems, including a novel construction for the finite field Kakeya conjecture. Gemini Deep Think (frontier LLM) verified it symbolically, and AlphaProof (math prover) formalized the proof in Lean for dimension 3.

You can read the full paper and see other obtained results here.

Throughout 2025, AI has made significant progress in advancing the frontier of mathematics. However, it’s essential to note that this is not a fully fledged AI mathematician, but rather a tool that real humans can leverage to make discoveries.

In other words, the role of humans is still fundamental to take AI in the right direction. Still, it’s nice to see that some of the most highly respected mathematicians on the planet, like “world’s smartest man” Terence Tao, actively acknowledge their utility and use them daily.

The AI startup FutureHouse has presented Kosmos, an advanced AI Scientist, now commercialized through their spinout Edison Scientific.

It utilizes a structured, continuously-updated world model to handle vast amounts of data (far exceeding typical LLM context limits), allowing it to read up to 1,500 scientific papers, generate 42,000 lines of code, and conduct coherent hypothesis testing in a single 12-hour run across fields like neuroscience, materials science, and clinical genetics.

Designed for transparency, every conclusion is traceable to specific code or literature sources. In beta testing with academic collaborators, Kosmos achieved seven validated discoveries: three reproducing unpublished findings and four novel contributions (e.g., linking SOD2 levels to cardiac fibrosis).

Importantly, it boasts a 79% reproducibility rate, with users estimating each run equates to six months of human research effort. However, it has caveats, such as 20% inaccuracy and a tendency to pursue irrelevant paths, positioning it as a powerful accelerator rather than a perfect tool.

In layman’s terms, this is what appears to be a very well-crafted agent orchestrator, a system that can handle and manage multiple different agents (LLMs, data analysis agents, literature search agents, among others), governed by what they call a world model that aligns their efforts in pursuit of a given objective.

This is loose cousin to Claude Research but for scientific discovery, which enables the system to go far beyond what a single research agent can think or search, but applied more broadly to scientific research.

Put simply, it’s a very smart multi-agent product, but the claims they make are outrageous; they literally say you can automate 6 months of work in hours, quite the bold statement, but we’ll see.

What I can say is that my prediction in late 2024 that AI would make its first discoveries in 2025 was spot on, if you’ll excuse this poor self-praise on my part.

Gemini is coming to Siri. Google’s upcoming frontier model family, Gemini 3, which is reportedly scheduled for this month (as soon as November 18th or earlier), will also be Siri’s backbone AI for its upcoming revamp, according to Reuters.

Interestingly, Apple appears to have inadvertently revealed the model’s size, which is estimated at 1.2 trillion parameters, and will (again, allegedly) pay up to $1 billion per year for the service.

The size feels “too small” for a Pro model, so it’s most likely Gemini 3 Flash, a model that, either way, should set all records compared to the current frontier line-up.

This would also be another significant victory for Google, which would secure a solid revenue stream of $1 billion, but I can’t help but feel underwhelmed by the number.

It seems that Google might have offered the service at an outrageous discount, trying to outcompete the other suspected alternatives, which are in much greater need of revenue.

And considering how low AI-heavy software services are for Google, this is more a branding victory than a meaningful source of profits (I think they’ll actually lose money on this one), which seems worth it for a company generating 400 times that number in revenues on an annualized basis.

ChatGPT has released a long-awaited feature (at least for me) that allows you to intervene in the model’s thinking and update your request without interrupting the model’s reasoning process.

As you can see above, I have added an update to focus on the larger form factor, which updates the reasoning chain to include the new requirement.

This is incredibly useful for someone like me, who interacts with reasoning models more often than with standard models. I’m also very prone to remembering things midway through the model’s reasoning, so this is simply a great feature.

As a caveat, it’s not yet supported on the desktop version (like many other features, such as connectors), which is not ideal for me, a staunch defender of desktop applications. However, the feature is truly worth the entire subscription if other providers don’t follow suit soon. To me, this is proof of the undeniable reality that the moat is the product, not the AI.

Moonshot AI, one of the top Chinese AI Labs, has released the reasoning version of its Kimi K2 model. The results are pretty staggering, especially when the model is allowed to utilize tools, setting multiple records across several benchmarks, even when compared to the top US models (GPT-5, Claude 4.5 Sonnet, or Grok 4).

You can try the model here, but you’ll have to fund your account first (it’s not free).

Kimi K2 is a unique model in its own right because the data pipeline used by the model differs significantly from those of other models.

Kimi K2 was trained not to generate chains of thought, but to interact in agentic settings as a means to facilitate the emergence of reasoning. In fact, they claim the model can execute up to 300 consecutive tool cools without human interference, an astronomical number even for today’s standards.

In layman’s terms, we usually get reasoning models by teaching them to break problems into smaller steps. Eventually, the model learns to approach problems this way, mimicking how humans tackle complex issues. Instead, Kimi K2 was trained by being exposed to environments and tasked to solve tasks, which led to the emergence of “reasoning behavior” but without chains of thought, and more focused on tool-calling. It’s unclear whether Kimi K2 Thinking is a mix of both approaches.

However, the takeaway for me is that we might be discussing a new ‘whale event’ similar to the disruption caused by the release of DeepSeek R1 back in March. I don’t think we’ve ever seen a Chinese model this good, to the point of being potentially considered the best on the planet.

The run may be short-lived, considering Gemini 3’s imminent release (and I’m pretty sure the model is worse than GPT-5 Pro or Gemini 2.5 Deep Think, even if it’s not an apples-to-apples comparison because these two have multiple models behind). Still, it represents clear proof that China has effectively closed the software gap.

Few technologies are as captivating to watch as diffusion LLMs. Now, the main AI Lab working on them, Inception, has released a new version they claim offers the best quality per unit of latency on the planet, at least compared to ‘mid-sized’ frontier models like Claude 4.5 Haiku or Gemini 2.5 Flash:

Naturally, their main feature is speed, which is hard to disagree with, as seen in the GIF above. The reason is that they ‘uncover’ the result instead of predicting it word by word. In other words, they work very similarly to image and video models, taking a blank slate and transforming it into the final solution through an iterative denoising process.

But what is diffusion?

I always like to use the analogy of sculpturing, where we “uncover” the final sculpture from a piece of marble or rock; the sculpture was there all along, we just erased the excess to uncover it, in clear inspiration of Michelangelo’s way of defining what sculpturing was.

Very cool technology. The question is whether being so much faster is truly instrumental for most coding tasks. Personally, I always prefer maximizing model quality and waiting longer than getting worse results faster.

But there’s certainly an appeal for software engineers who spend hours on end coding; balancing the speed of diffusion LLMs for boilerplating and more coarse coding, and then refining with superior autoregressive models like GPT-5 Codex.