Hey, my friend,
How were the last two weeks for you?
I just arrived in Texas, and spent some time making sure I don’t write in a tent but in a flat with an air con (very important for Texas, of course, but maybe equally important to those of you in the UK these days).
There have been fewer AI headlines in the last couple of weeks. That said, the executives aren’t any less busy trying to use their AI to dominate the world.
First of all, it’s a release season for OpenAI, apparently, Astra (or so-called GPT 6), Jalapeño (not the chill, but a chip, and no, not those chips).
Then I’m also covering the marriage between Nvidia and Hugging Face; Is SaaS dying, yet? Apple’s new product (isn’t iPhone Dou, but Ternus).
Last but not least, Anthopic’s attempt at unifying the AI to physical standard, think the USB for AI.
Shall we?
OpenAI has been really busy.
Vertical control
So, OpenAI delivered Jalapeño to handle AI requests faster and use less electricity (up to 1.9× more inference work per watt), which, in theory, makes every ChatGPT answer cheaper to deliver.
As Sam Altman tweeted, “We made a chip and it is fast,”
Nearly 2 times more inference per watt! Does this mean we can see the relief of the energy issue?
If you’re excited about this news, I’m stopping you right here.
Just think about this: what’s the chance that a company that focuses on an AI model suddenly beats one that spent decades making chips? Yes, OpenAI has the help from Boardcom, but there’s also a reason why Boardcom isn’t leading the industry.
So, a couple of catches:
Jalapeño isn’t 2 times better or cheaper than Blackwall (the world's most advanced GPU by Nvidia) because we still need to take into account factors like manufacturing cost, depreciation, networking, or even OpenAI’s actual cost per token.
Jalapeño doesn’t beat Blackwall at everything; in fact, far from it. If Blackwall is a Swiss knife, then Jalapeño is like a steak knife, for steak only. Blackwall can be widely used for training and inference, whereas Jalapeño is for inference only (for example, dedicated to running the ChatGPT query you sent).
In my piece about AI companies running into physics, I argued that the cleanest profit pool keeps appearing below the model layer: chips, data centers, and power. While the model company shows the bling bling, the margin, however, stays with the infrastructure suppliers.
When ChatGPT becomes more useful, people use it more; that’s true. However, unlike traditional software, AI use does not approach zero marginal cost; instead, it quickly becomes more expensive.
I’m not saying there’s no way to keep this under control; there's more effective practice every day, but it’s still much more expensive to run than software.
Therefore, Jalapeño is a sensible first step for OpenAI to take control of its bills. Though it’s not versatile, it’s still a better option for OpenAI to meet the recurring billing needs of its millions of customers vs. relying on the external option.
With this update, OpenAI can expect better margins across its ad revenue, subscriptions, and enterprise API queries.
All rainbows and unicorns that Jalapeño may lower OpenAI’s costs. Yet, it doesn’t make OpenAI a chip company overnight.
For now, it’s an attempt to make ChatGPT and Codex work without handing quite so much money to Nvidia. Which brings us to the next story, because OpenAI’s balance sheet isn’t pretty.
An impressive number, but still one that missed the mark
Also, OpenAI celebrated that ChatGPT Ads reached a $1 billion ARR in less than 200 days.
Hooray?
Not too fast.
I believe there should be a regulation enforcing business announcements like this to be put in context, and that CEOs should be held accountable for all their open promises.
First, early this year, Axios reported that OpenAI told investors to expect $2.5 billion in total advertising revenue in 2026. Its newly announced $1 billion annualized pace is therefore well below that earlier full-year target.
Although OpenAI could still accelerate sharply as it expands and opens self-serve buying to more advertisers.
But what’s more important here… is that Altman preferred not to profit from the ads, at least that was what he said last year to Harvard’s elites:
I kind of think of ads as, like, a last resort for us as a business model. I would do it if it meant that was the only way to get everybody in the world access to great services. But if we can find something that doesn’t do that, I’d prefer that.
Regardless, if you do want to find out its true profitability in ad, you should ask these:
How much advertising revenue has OpenAI actually collected? Does it come from a few big advertisers?
Are advertisers renewing their contracts?
What computing power is needed to cover the conversations surrounding those ads?
Meanwhile, the relentless OpenAI’s pursue to profitability doesn’t just stop here. It’s also fighting over whether it must pay the publishers whose work was the cornerstone of its products.
Something OpenAI would rather not pay for, even when it’s illegal
If Jalapeño is one input that OpenAI wants to make cheaper, then the training material is another.
The Times did not wake up one morning and sue because ChatGPT sounded vaguely newspaperish. It started building the case in 2023, after tons of unambiguous evidence that AI chatbots both absorb vast amounts of paid journalism and answer readers’ questions without sending them back to the publisher.
To keep it simple, OpenAI, on the stand, says this is entirely fair use of knowledge, while NYT says otherwise.
The latest is that the Trump Justice Department has officially sided with OpenAI.
A boring 20-page document keeps repeating the term ‘fair use’.
The DOJ’s defense is basically: OpenAI training on the whole internet is like teenage Joan Didion (I wonder they picked her because she couldn’t really fend for herself) typing out Hemingway in her bedroom.
mm…
One was a teenager learning to write. The other is a company with a roughly $900 billion market value.
Boo hoo! Astra is so dangerous!
Before we move on to other AI news, here’s the last one about OpenAI.
There is a new fashion in AI: the CEOs launch their models by explaining, with grave concern, why it may be dangerous!
While these claims are presented as a warning. They are also, somehow, the sales pitch. “Please take our safety measures seriously” is increasingly AI-industry code for “that’s how far ahead we are,” and “please update your estimate of our valuation!“
Meet GPT-6 Astra, OpenAI’s latest model.
According to OpenAI president Greg Brockman, Astra is an AGI, and the company also claims that Astra is the first model to reach its “Critical” cybersecurity threshold.
Given the right access, OpenAI says it can find unknown vulnerabilities and turn them into working exploits without a human holding hands.
Apparently, that is alarming enough to require special safeguards but reassuring enough to release as a product.
That said, the safety concern is not entirely theoretical. If you still remember, the instance in which OpenAI agents broke into Hugging Face, I covered it here.
That agents may not be the exact Astra model, and perhaps the new safeguards work perfectly. We just don’t know yet, what we do know is that danger has become remarkably useful marketing.
It is the first industry where the warning label doubles as the product brochure.
That said, is AI dangerous enough to threaten the whole SaaS industry?
Salesforce earnings
Some believed the answer to this one is a yes since 2024.
I still remember when I was invited to an event full of SaaS CEOs to talk about AI’s progress and how they can implement AI in their firms.
I could see how the host and some of the guests were expecting a magical "pulling a rabbit out of my hat" demo. One of them even firmly believes that AI will replace all SaaS soon!
I mean, yes, the person will eventually be right. Just like many people say the AI bubble is popping, to an extent, they are right; it’s expanding, and sure, at some point, it pops.
Two years later, has SaaS died yet?
Salesforce reported $11.35 billion in revenue for the last quarter, up 11% YoY. It also implies 11%–12% growth for the year ending next January.
The software panic, to recap, was simple.
AI would turn every expensive subscription product into a feature. People start questioning why pay for Salesforce, ServiceNow, Workday, or Asana if a competent model can fill out the form, summarize the meeting, or answer the employee's question?
Yes, plenty of SaaS companies spent years charging separately for small conveniences that a general-purpose model can now approximate.
But keep in mind, enterprise software is not merely a collection of buttons.
It is where a company’s customer data, employee permissions, and institutional habits live. Replacing that infrastructure is nothing like canceling Netflix; it's more like replacing the plumbing while your partner is using the bathroom (probably also depends on how much you hate each other).
The market has been confused.
Apart from Salesforce, Workday’s earnings offer the same market feedback to those investors or analysts who don’t understand the enterprise software market and don’t listen.
Workday’s fiscal second-quarter revenue rose 12.8% to $2.65 billion. Plain and simple, their customers haven’t stopped paying for software.
As I keep emphasizing, the incumbents will be the ultimate beneficiaries of this LLM’s improvements. There will be pain, yes, but they will also soon learn what value + AI can be quickly offered to existing clients.
Or let me ask you this question:
Is it better to build one AI feature from the ground up and look for clients to buy it, or is it easier to win when you only need to make some integration and resell it to your existing customers?
NVIDIA bought Hugging Face
I have been writing about vertical integration a lot lately.
Musk bought Cursor because xAI had the compute but lacked developer distribution. Now, Nvidia is solving a similar but larger-scale problem: it also wants more influence over what developers choose to run on it.
The $12.9 billion deal to buy Hugging Face is the clearest evidence yet.
Think of Hugging Face as a marketplace.
Millions of models live there. Hugging Face sits between someone releasing a model and a developer putting that model to work.
Nvidia wants to own that choke point.
Here’s what Jensen Huang pays for: developer habits, model distribution, and even more influence before anyone decides which hardware to use.
It is the same vertical-integration logic I discussed in my piece about the real AI bottleneck.
There is a defensive side, too. Nvidia’s biggest customers are building their own chips because they are tired of paying Nvidia prices.
At the same time, Nvidia has invested heavily in the companies and data centers buying its hardware. If all that expensive capacity ever needs more customers, owning one of AI’s largest distribution platforms could be rather helpful.
The challenge for Nvidia after the acquisition is to maintain Hugging Face’s independence (or the facade of it). Its value comes partly from being Switzerland, ie., developers can choose freely without pledging loyalty to one company.
Nvidia started the AI boom by selling picks and shovels. Now it is buying the map developers use to find the mine.
Can’t wait to see how others might counter this move.
The USB for AI?
Do you remember what it was like before USB?
Adding a printer could become a family project. You needed the right cable, the right port, the right driver, perhaps several reboots.
Then USB changed everything.
And Anthropic is trying to replicate this moment.
Today, getting an AI agent to operate one of those machines can require a lot of custom work. Anthropic calls it Model Hardware Standard, or MHS, which is meant to give machines such as microscopes, lasers, and robots a common way to communicate with AI.
So instead of paying engineers to build tailored comms for every robot, camera, and factory machine, it would be much cheaper and also speed up the final product delivery.
However, it also means that for some companies, proprietary connectors and painful integrations may no longer be much of a moat, and the focus has to return to the product itself.
Well, it’s probably too early to say any of these; this is still a theory by Anthropic. But something worth keeping an eye (or both) on.
Apple’s next chapter
Is Thursday’s Apple event a launch of the latest Apple products or of its new CEO?
Tim Cook, from 2011 to 2026, grew Apple’s annual profit fourfold to $112 billion, and its market value from about $350 billion to more than $4 trillion.
Cook leaves behind a much larger company, an extremely profitable iPhone business, and no disaster for his successor to fix.
That is what makes the handoff difficult, especially for Ternus. He has inherited a successful company whose next chapter is unclear.
Ternus first major launch as Apple CEO was another iPhone, now with a hinge.
But the phone market is mature. Foldable or not, this isn’t likely going to be the product that extends Apple’s lifeline. Instead, Apple needs another category that expands its market without simply moving customers from one expensive Apple device to another.
AI could help, I’m not saying that Apple needs to beat OpenAI or Google at building the smartest model. In Is Apple the Only One Getting AI Right?, I already argued that Apple’s advantage is putting useful intelligence directly inside the device. It controls the chip, the OS, and the customer relationship. Not to mention, local AI also reduces cloud costs and keeps more personal data private.
The most obvious product that lives well with AI is the Mac Mini.
But this isn’t enough, because this is mostly a developers’ market.
To gain a larger market share, Ternus now has to turn that advantage into a reason to buy Apple hardware. Otherwise, AI becomes another free software feature that makes an existing iPhone slightly better.
I guess you are the same as me, keen to see how Ternus is giving it a future that is more than a polished, increasingly expensive version of its past.
Gates’ AI warning.
Speaking of the dangers imposed by AI.
In a long essay, Gates argued that AI brings risks serious enough that he would support a global slowdown in progress, that is, if anyone could produce a credible way to make one work.
That is a large “if.”
The companies speeding it up, however, keep pointing to the same defense: it is becoming useful. For example, Jensen Huang says
AI has reached its inflection point. It’s doing useful work.
and that demand is accelerating because “tokens are productive and profitable.”
Or in the latest earnings, how Microsoft points to growing cloud demand, and Alphabet says AI is helping sell more cloud capacity and improve its advertising machine.
Some of the risks Gates raises should sound familiar to you.
I spent much of 2024 and 25’ arguing against this sci-fi claim: that chatbots were already ready to replace large numbers of workers, especially junior white-collar workers.
When the models were too obviously error-prone, too brittle, and too dependent on human supervision. Receiving a wrong answer from a junior is annoying. Sending out 10,000 incorrect answers within minutes as a company is catastrophic.
I still do not think “AI can do a task” automatically means “AI replaced a job.”
Those who said that have no idea how long it takes to change one single business process in a hundred or even thousand-person company.
In this AI replace job theory, more often than not, the business has no workflow to adopt, no accountability for errors is needed, and the AI is like an item by a senior high school teacher that always sees itself isolated in a perfect spherical vacuum condition.
That said, I get Gates’ concern, not that AI wakes up one morning and fires everyone, but it is that it is quietly improving, to an extent that it’s no longer that error prompt… Then what?
A good enough tireless machine that can do a piece of work with no error.
An AI system only needs to handle a useful share of the repetitive work with an error rate low enough that we no longer need a verifier and another verifier, then the question ‘will AI replace some jobs?’ slowly becomes yes.
When the answer slowly moves towards yes, the effect may not arrive as a spectacular wave of layoffs to make millions feel the pain.
Though six months (since I last criticized this kind of claim) is not enough time for AI to be error-free, you also shouldn’t need to doubt that this is still a messy labor market, a high-rate environment by post-2008 standards, uneven hiring across sectors, and companies eager to sound “AI-first” whether or not they have actually changed anything.
But all that said, I believe the direction of travel matters.
It will be too late to only think about the problem when it becomes obvious.