
THEWHITEBOX
TLDR;
Welcome back! This week we cover Anthropic’s S1, OpenAI's latest AI-enabled discoveries, new models like Gemini Carbon, and earnings data that reveals the industry's financial state.
Enjoy!

THEWHITEBOX
Anthropic’s S1 has a lot of surprises
I’ve provided a detailed rundown of everything you need to know about this IPO and more in a recent Medium article you can read for free here.
We finally have access to Anthropic's financials as it prepares for an imminent IPO. And, well, as expected, it’s a lot.
Anthropic reported $16.3 billion in revenue for the first half of 2026, up from $1.2 billion a year earlier and more than three times its revenue for all of 2025, according to its IPO registration filing. Second-quarter revenue reached $11.6 billion, compared with $4.7 billion in the first quarter. Add the leaked $16.8 billion from the third quarter, and Anthropic’s revenues are on an unprecedented growth trajectory.
However, the company recorded a net loss of $80.7 billion in the first half. This sounds outrageous, but it’s kind of a red herring because $72.4 billion of that charge comes from changes in the fair value of convertible notes and other financing instruments. In simple terms, Anthropic issued convertible notes, loans that may or may not carry interest, that include a right to a certain number of shares at a very discounted rate, and the associated shares have increased in value by a lot.
Thus, the $72.4 billion “loss” refers to the growth in value of those shares since the agreement was made (the principal amount is $6-ish billion), which actually speaks more to Anthropic’s incredible valuation growth and less to something investors should be panicking about, especially considering this loss had no impact on cash (it’s a paper loss).
Anthropic posted an operating loss of $693 million for the six-month period, but reported $974 million in operating income in the second quarter, an 8% operating margin, the first time in its history to achieve operating profitability. It also generated $3.9 billion in cash from operations in the first half, compared with $1.5 billion used in the same period of 2025.
Despite the alleged unprofitability of the AI trade, this looks pretty profitable and cash-flow generating, right?
Well, hold your horses for one second. But before things turn south, which they do, let’s cover the other positives.
As of September 30, Anthropic had more than 11,900 customers generating at least $100,000 each in revenue over the preceding 12 months, including more than 1,900 generating at least $1 million, and 250 spending $10 million a year or more.
And most important of all, the revenue/compute ratio, the most important metric in AI that measures how efficiently a business converts compute spend into revenue, showed a 2.08x increase in Q2 (although that number has since dropped to 1.7x in Q3).

This means Anthropic generated $2 for every $1 of compute spend. Naturally, one must include other spending to see if a company actually makes money (e.g., salaries, marketing, non-operational expenses, etc.), but it’s a great sign.
Nonetheless, Anthropic’s margins aren’t even better simply because they spend almost 80% of compute on training, meaning that their inference margins, the ratio of what they charge for their AI models relative to what they spend to serve those customers, are close to 90%.
As I always say, AI is not a profit problem; it’s a cash flow problem.
But here are the good things to say about this release.
The biggest problem with Anthropic is not its present; it’s its future. In clear view, potentially up to a trillion dollars in spending over the next two decades just on compute, an amount of money they obviously don’t have right now.
Therefore, Anthropic’s future hinges almost entirely on its ability to dramatically raise its revenues. For context, Anthropic’s compute spending, based on already-agreed commitments, could rise to almost $12 billion in 2028, and that could still fall far short of the eventual value.
At its current revenue/compute ratio, that means Anthropic will need $24 billion in revenue per month, around four times its current last month (allegedly) in a little over a year to support that while maintaining compute margins, a rate of almost $300 billion/year in revenues, almost as much as NVIDIA’s revenues today (roughly $300 billion/year).

Is that impossible? No. Unlikely? Yes.
Worse, approximately half of the already-known commitments are non-cancelable (around $430 billion), meaning Anthropic can’t simply back out unless, well, they default.
TheWhiteBox’s takeaway:
My Medium article is a 4,000-word rant about the entire thing, in case you really want to dive deep.
But my main takeaway is that this company has presented financials showing a reasonably healthy state that will change drastically over the next months and years. Or, more explicitly, investors will be buying a company radically different from the one they’ll own a year from now.
This isn’t nothing new; Amazon went public decades before it turned a profit.
But investors knew what they were getting into. I’m not sure future Anthropic investors know what they are getting into.
WORLD MODELS
Google Marvels the World with Genie 4 (Again)
In a world where AI is basically all large language models, at least in the eyes of most of society, one particular model family from Google, Genie, continues to remind us that, in fact, there’s way more to AI than LLMs, and that world models could be as important, or even more, in the future.
And its new version, Genie 4, is about to blow your mind completely. But what is Genie?
In short, Genie is a world model, an AI that has been trained to predict what will happen next and can use that ability to not just produce AI-generated videos, but make them interactive.
In other words, the person creating the video not only dictates the theme, but can control what happens in every single next frame by interacting with the video. If you create a world of a boat across the sea and you ask Genie to turn the boat left, the boat turns, and that affects everything that occurs after that action, meaning the AI model’s future generated frames have to be consistent with the decision the user made.
Put simply, it’s not “create a video of a boat sailing across the sea,” but “create a simulation of a sailing boat, and I’ll get to control, in real time, what the boat does.” So Genie 4 generates the new frames and the next events in real time as you make decisions.
I like to think I’m a good explainer, but in this case I highly encourage you to go to the link and watch the demos and then come back.
Pretty incredible, right? (I’m assuming you did, in fact, press the link, which you should).
Where I can contribute a little more is regarding how this is done. This is proprietary technology, but we know the foundation because the first-ever Genie was open-sourced back in 2024.
After having watched Genie 4’s videos, you’ll immediately grasp the extent of the progress Google has made with its Genie family since the first version, which you can see below:

So, how does it work?
All AI models, especially generative ones (the ones that produce tangible output like text or video), are prediction engines: you give them an input, and they predict something in return.
LLMs predict the next token in the sequence, usually a word. Image models can also predict the next token or tokens, usually pixel patches. And the same applies to video generation.
So it’s always a prediction based on an input. The difference here is that the input is partly provided at run time.
If we think of a standard video generation model, you ask for something like “a busy street in downtown Manhattan,” and the model will generate the entire sequence. In that case, the model is concerned with two things:
Ensure the entire sequence responds to your input request
The different frames in the video are coherent with each other. In other words, each frame is generated largely based on what the previous frames showed; otherwise, the video becomes inconsistent (e.g., a ball of a certain size changes size in the next frame).
What makes the Genie family of models particularly incredible is that to generate the next frame there’s a third variable: user actions. The next frame’s generation is conditioned not only on the original request that defined the scene and the previous frames to ensure consistency, but also on what the user decides happens next in the moment.
The user can make the character jump, the boat turn, the airplane nosedive… the user controls both the original theme and the infinite futures that world can evolve into.
That is, Genie 4 doesn’t create videos; it creates simulations.
As for its applicability, beyond the obvious (e.g., video games), the more latent element is the “intelligence” this model needs to understand our world.
It takes no genius to realize that a model that is trained to predict what happens next in open worlds with infinite things to account for (e.g., balls in the air need to fall down, if wind blows in one direction, tree leaves can’t blow in the opposite direction, and so on), that model is equipped with a level of spatial and world intelligence that an LLM trained mostly on text should not have.
Many researchers have theorized that world models are, in fact, a better foundation for intelligence than LLMs trained mostly on text and code, because they are exposed to the rules of physics, among other things.
This has led many, including Google DeepMind’s Chairman and Nobel Prize winner, Demis Hassabis, to theorize that the future of AI models will be a mix between Genie-type models working alongside LLMs to provide the required degree of “intelligence” depending on the task at hand.
This view hasn't materialized and isn’t close, but it’s hard to fathom a future where LLMs can reach the uncanny level of simulation generation Genie achieves.
Either way, if you thought AI ended with LLMs, think again.
DRONE TECHNOLOGY
A Drone to do… exactly what you say?

Purdue University’s IDEAS Lab has demonstrated a drone following a person through a forest, with the lab saying it navigates without GPS, a pre-existing map, or cloud computing. Lab director Aniket Bera posted a 44-second video of the flight on October 23.
According to Bera’s description of the demonstration, the drone uses onboard AI to process information and predict the person’s path when trees temporarily block its view. The video also shows a map generated using LiDAR, which uses laser measurements to detect the surrounding environment.
Separately, Purdue announced a five-year, $1.5 million agreement with the U.S. Army in November 2025 for research into aerial and ground robots that can map and navigate terrain without GPS. The demonstration has not been publicly linked to that funding.
TheWhiteBox’s takeaway:
The videos are pretty incredible, but I find it hilarious that they call this a system to rescue people in the forest, because we all know that the first thing you do when you need rescuing in the forest is to run away from the drone that is actively trying to rescue you, right?
It’s palpably obvious, especially if you have the slightest idea of what is taking place in Ukraine, that this is a human-targeting drone technology.
RESEARCH
How Much do Agents Influence Experts?
Professional analysts followed their AI agent’s randomly assigned preference in 86% of key analytical decisions in an Anthropic study published on October 27. Researchers said they would have expected agreement around 50% of the time if the agent’s preference had no influence.
The experiment asked 50 analysts to use Claude Code to assess whether a fictional farming cooperative should expand an advisory program. For each decision, Anthropic’s researchers secretly instructed the agent to favor one of two reasonable options, selected by coin flip. The agent was told to make an honest case for its preference and change its mind if the evidence warranted it.
The choices included how to measure productivity, which farms to study, and how to weigh their results. The agent’s preference also influenced the final recommendation: 54% of analysts recommended expanding the program when their agent favored expansion, compared with 36% when it favored discontinuing it.
Afterward, 77% of analysts repeated an argument made by their agent in at least one explanation, sometimes nearly word for word, while 15% credited the agent in any answer. Among the 48 who completed the survey, 52% said the final recommendation reflected their own judgment more than the agent’s.
TheWhiteBox’s takeaway:
This isn’t surprising to me. I can tell AIs shape some of my opinions when I discuss them with them. In theory, that’s not necessarily bad, since I’m using the agent to gather information, which enriches my context and helps me make better decisions.
What concerns me is that if we end up with just a bunch of model providers, the companies behind them, will have incredible power to shape public opinion.
Now you have plenty more reasons to ensure open-source wins, and we end up with a diverse lineup of models with different views instead of the alternative: whatever the owners of that model want society to accept or believe.
MATHEMATICS RESEARCH
Another Day, Another Science Discovery
OpenAI says its AI models have produced three research results in mathematics, computer science and physics, including a claimed disproof of the Exponential Time Hypothesis, a widely held assumption about how long certain computational problems take to solve. The company announced the results on October 26 and is working with outside researchers to examine and explain them.
In mathematics, OpenAI says its models produced a proof of the local smoothing conjecture in every dimension, which would also resolve the Kakeya and Stein restriction conjectures.
In computer science, the company says its system disproved the Exponential Time Hypothesis, or ETH.
Both results were formalized in Lean, a tool that checks mathematical proofs, and OpenAI has published the underlying AI outputs.
OpenAI also claims its models produced mathematical existence proofs for quantum chromodynamics, the theory describing the strong force, and for a simpler Yang–Mills theory. The latter addresses one part of a Millennium Prize problem but does not resolve the separate “mass gap” portion.
The physics materials have not been released publicly. OpenAI says an external group led by physicist David Gross will decide how to release them. The Advisory Group on Mathematics and Artificial Intelligence has called for their immediate publication, arguing that researchers should have equal access to examine the claims.
TheWhiteBox’s takeaway:
The ETH is particularly exciting/concerning. Essentially, it’s a famous conjecture in computer science that certain complex problems have no clever shortcuts. To solve these problems, a computer must use "brute force"; guessing and checking a list of combinations that explodes in size with every new variable.
As the problem grows, it quickly becomes so overwhelming that even a supercomputer would take billions of years to crack it.
Okay, so what? Well, because modern cryptography that secures our digital systems rests on the idea that ETH is true and that there are, in fact, no shortcuts.
For example, modern encryption turns our digital communications into ciphertext, so if someone who’s not the intended recipient intercepts the message, it will read as gibberish. The recipient has the secret key to decrypt the message, so we’re good; without the secret, the eavesdropper would take forever to find the key that decrypts the message.
But if these shortcuts exist and are found, then all modern communications are at risk. And while OpenAI hasn’t presented any shortcuts, they have proven that the conjecture, the assumption that there are no shortcuts, is false, so the chance that these shortcuts exist is not zero.

MEMORY
SK Hynix Has Blowout Quarter
SK Hynix reported record third-quarter revenue of $78.9 billion and operating profit of $62.8 billion, as rising memory prices and sales to AI data centers lifted results. Revenue increased 328% from a year earlier and 32% from the previous quarter, while operating profit rose 632% and 38%, respectively, according to its preliminary earnings release.
The company’s operating margin reached a record 80% (for reference, Anthropic’s historic operating margin in Q2 was just 8%), and net profit was $30.0 billion, up 216% year over year but down 58% from the second quarter, a number that feels outrageously low relative to operating profit.
And the answer was related to investments. On its earnings call, SK hynix said it recorded approximately $24.6 billion in net non-operating expenses, including $17.3 billion in investment-asset valuation losses and $8.8 billion in net foreign-exchange losses. Other non-operating items offset part of those losses.
For the fourth quarter, SK hynix expects HBM4, its latest generation of high-bandwidth memory for AI processors, to surpass HBM3E in both shipments and revenue share, a good sign after months of concern about the yield (% of functioning chips) of the upcoming generation.
At quarter-end, SK hynix held approximately $86.3 billion in cash and cash equivalents and $11.9 billion in interest-bearing debt. It plans approximately $75.4 billion in shareholder returns for 2026, including at least $22.6 billion in cash dividends and at least $45.2 billion in share repurchases and cancellations.h3
TheWhiteBox’s takeaway:
Another historical quarter for the memory companies. But the question, as always, is whether these revenues are sustainable. The fear with memory companies is whether these historical results will remain for the foreseeable future.
As long as investors believe results are not sustainable, these companies trade well below standard price-to-earnings multiples. Using SK hynix’s October 28 Korean closing price, I get an approximate trailing P/E of 5.7× and a 2027 forward P/E of 3.1×, the latter based on projected future revenues.
In layman’s terms, the company’s valuation is three ~6 times its profits over the last 12 months and only 3 times next year’s projected profits. For reference, SK hynix’s trailing multiple is about 74% lower, and its forward multiple is about 84% lower, than the average S&P500 company.
If SK hynix traded at the S&P 500’s earnings multiples, its implied market capitalization would be approximately $3.2 trillion using trailing earnings, or $5.0 trillion using estimated 2027 earnings, instead of the current ~$800 billion.
The only reason it doesn't is cyclicality: fears that Hynix’s revenues and profits are cyclical (meaning they have great moments followed by serious downturns—less than four years ago Micron had negative gross margins), which has always been the case for memory companies.
However, I don’t think that is true anymore. I’ve made my case multiple times that memory is no longer cyclical; it is an essential component of modern AI technology. AI is a fundamentally memory-centric paradigm, so as long as there’s demand for AI hardware, demand for memory will be enormous.
And I don’t seem to be the only one who thinks this; chip buyers think so too. Chinese cloud providers Alibaba, Tencent and ByteDance are seeking long-term supplies of server DRAM from SK hynix and Samsung, Maeil Business Newspaper reported, citing industry sources. Some customers have requested contracts lasting until 2035, beyond the three-to-five-year terms the publication says are typical.
If ten-year agreements sound short-cyclical to you, that's because memory’s previous reality was short-cyclical, and I don’t know what else to tell you.
Of course, the pushback is that we can’t guarantee that the demand for AI hardware will last forever, and I agree. But then, that applies just as much to NVIDIA, AMD, or any other compute chip company, companies that trade as if they weren’t cyclical.
Why are Hynix and other memory suppliers being treated differently? I’m sorry, but that makes no sense.
It’s one or the other; the market can’t have both.
PUBLIC MARKETS
Samsung Sets Another Record… And Yet Falls
Like Hynix, Samsung Electronics also reported record-hitting earnings, with third-quarter revenue of about $146 billion, up 127% year over year, and operating profit of $80.5 billion, up 783%. Net profit rose 515% to $56.2 billion but fell short of analysts’ expectations.
The semiconductor division generated about $80 billion in operating profit, equivalent to 99% of Samsung’s total, making Samsung the most profitable company in the world by operating profit. In simple terms, no business on the planet has a more profitable business model than Samsung.
Excitingly, the company said sales of its latest HBM4 memory chips more than tripled from the previous quarter and that its high-bandwidth memory supply for 2027 was sold out.
Despite the strong results, Samsung shares fell as much as 6.2% in Thursday morning trading. Analysts cited by Reuters pointed to expectations for more detail on shareholder returns and a less bullish outlook for memory shortages than rival SK Hynix’s.
TheWhiteBox’s takeaway:
My thoughts are the same as what I wrote above for Hynix. The only nuance here is Samsung’s foundry business, which produces advanced chips in rivalry to the king of that segment: TSMC.
Sadly, Samsung did not disclose a separate revenue or profit figure for the foundry in its earnings release. Reuters reported that analysts believed the business remained loss-making, despite the improvement. So we can say its results got better, but we can’t tell from Samsung’s release how much revenue it generated or how large its loss was.
The foundry business remains Samsung's big ‘what if.’ If they deliver, they would be a primary player in the two biggest markets in AI today: memory and chip manufacturing.
PUBLIC MATTERS
Anthropic Solidifies Relationship with the Government
Anthropic and the U.S. General Services Administration have expanded their agreement to give government buyers a 40% discount on Claude model usage. Agencies will pay for what they use, with no minimum purchase required.
The discount covers Claude Enterprise, Claude for Government and the Claude API, and extends to federal, state, local, tribal and territorial governments, Anthropic said. It is available for direct purchases and through GSA’s procurement schedule; Anthropic said it would extend to purchases through cloud providers in the coming months.
The agreement comes ahead of the October 31 expiration of an earlier federal offer that provided access to Claude Enterprise and Claude for Government for $1 per user. Anthropic estimates the new discounts will save the government $1 billion, without specifying a time frame or the usage assumptions behind that estimate.
Anthropic said approximately 500 federal, state and local government organizations use Claude. The company also plans to launch a program in the coming weeks to help agencies implement the technology.
TheWhiteBox’s takeaway:
Good news for Anthropic shareholders ahead of its IPO: the relationship with the government doesn’t seem as broken as it did earlier this year.
However, not everything seems solved.
The agreement comes amid Anthropic’s continuing dispute with the Department of War, which designated the company a supply-chain risk after disagreements over military uses of Claude. A federal appeals court upheld the designation in September.
The GSA announcement does not say whether the department can purchase Claude under the new agreement, which does not itself lift the defense restrictions.
EARNINGS SEASON
Microsoft & Google Earnings Show Very Different Outcomes
Microsoft shares rose more than 5% in after-hours trading after the company reported stronger-than-expected cloud revenue and issued an upbeat outlook for its current quarter, Reuters reported.
The company’s fiscal first-quarter revenue rose 19% year over year to $92.3 billion, while operating income increased 21% to $45.9 billion (a little bit more than half what Samsung generated, showing how incredibly profitable Samsung is).
Net income rose 32% to $36.6 billion, or $4.92 per share. Adjusted earnings, which exclude the impact of Microsoft’s OpenAI investments, were $4.89 per share, above the roughly $4.73 analysts expected.
Azure, Microsoft’s cloud business, revenue grew 48% to $33.17 billion, exceeding estimates of $32.84 billion. For the current quarter, Microsoft forecast Azure growth of 49% to 50% and total revenue of $96 billion to $97.1 billion, both above analysts’ expectations.
Quarterly capital spending reached $45.2 billion, below Microsoft’s $50 billion projection, partly because third-party capacity deliveries were delayed. Free cash flow was $18.6 billion, compared with an analyst estimate of $8.54 billion.
Alphabet, Google’s parent company, posted a very different result.
Google shares fell more than 10% in after-hours trading after the Google parent reported third-quarter results that missed revenue expectations, despite 25% revenue growth and a doubling of Google Cloud sales.
Revenue reached $127.6 billion, below the $128.2 billion analysts expected, according to LSEG data cited by Reuters.
Operating income rose 30% to $40.6 billion (half of Samsung’s), but net income available to common shareholders fell to $26.2 billion from $35 billion a year earlier.
The positive note was the cloud business. Google Cloud revenue rose 103% to $30.7 billion, with Alphabet saying sales of its TPU systems contributed meaningfully. Search and other advertising revenue rose 12% to $63.4 billion, while YouTube advertising revenue increased 9% to $11.2 billion. Both advertising figures fell short of some analysts’ expectations.
Alphabet recorded negative free cash flow of $9.1 billion as capital spending increased, meaning the company, for a second consecutive quarter, spent more than it made in cash.
And to make matters more complicated, Chief Financial Officer Anat Ashkenazi said the company still expects capital expenditures to increase significantly in 2027, with more detail to come on its fourth-quarter earnings call.
TheWhiteBox’s takeaway:
I can’t say much about Microsoft beyond the fact that it’s been a while since it became the most cautious of the bunch on spending. Markets punished them for that early in the year, but as the overall trade turned south in late July, they were handsomely rewarded.
In AI, we measure cautiousness by free cash flow, which remains very positive for Microsoft, with almost $20 billion in cash generated last quarter, and less so for companies like Google, which posted a negative $9 billion after spending $50 billion on AI investments in three months.
This chart alone could easily explain how investors reacted.

But Google’s “situationship” has another interesting fact: the disproportionate amount of non-operating net income in its results. For instance, in Q2, Google presented $112 billion in net income, of which 70% was non-operating profit, meaning only 30% came from Google’s own business and the rest came from places like investments (e.g., SpaceX).
I’m as bullish as they get on Google, but that wasn't healthy, and as soon as the non-operating profit bonanza faltered, the stock paid the price.


EDUCATION
Are These Tools the Future of Education?
A new tool from the guys at HeyClicky offers one of the coolest AI experiences I’ve ever seen. The idea, which I can’t fully explain using words, is one I invite you to see with your own eyes in this video, but it’s a process in which an AI model generates visualizations on the fly and renders them on top of the user’s screen, helping the person understand the concepts better.
To do this, the system runs OpenAI’s Sol 6.1 Ultrafast, a version of the model that generates far more tokens by running users in small groups, which skyrockets interactivity (speed per user) because the chips have to serve far fewer users.
The model generates on-screen renders in real time at acceptable speeds and makes them interactive based on real-time user instructions.
TheWhiteBox’s takeaway:
It’s really hard for me not to envision tools like this as the future of education: offering fully interactive, endless visualizations, schematics, and examples to nurture the student's mind.
ROBOTICS
Unitree's new Kung Fu robot
Unitree Robotics has introduced the H2-X, a humanoid robot presented in a promotional video as a “Kung Fu Legend Guardian” for security and rescue tasks. The impressive October 21 video shows the robot performing martial arts movements and acting out scenarios involving a knife-wielding attacker and a person trapped beneath a car.
The footage also shows the robot landing after a drop from a platform marked four meters high. The rescue scenes used props, and Unitree has not said whether the actions were performed autonomously, remotely controlled, or preprogrammed, according to a review of the video.
Unitree’s H2-X product page now lists the robot at $25,400 before tax and shipping. The company says it stands approximately 1.7 meters tall, weighs at least 65 kilograms, and has 31 joints. It lists approximately eight hours of battery life both while stationary and during continuous walking; the promotional video specifically described the eight-hour figure as standing time.
TheWhiteBox’s takeaway:
Once again, Unitree has put its humanoid’s physical capabilities on display. But there’s an important distinction between a robot performing impressive movements in a demonstration and one that can understand its surroundings and reliably carry out useful tasks on its own. Unitree hasn’t said how it controlled the H2-X’s movements in this video.
US companies such as Figure AI have placed considerably more emphasis on that second challenge, demonstrating robots using AI to perform tasks such as loading a dishwasher, while Chinese labs place more focus on the physical side.
Who’s right? Still too early to tell, but future humanoids will need both.
MODELS
Google Accelerates with Gemini 4 Carbon

Google announced Gemini 4 Carbon, a new AI model focused on cybersecurity and complex software engineering, two weeks after releasing Gemini 4 Argon. Access will initially be limited to a subset of cyber defenders, with a broader release planned after further testing.
Google says Carbon scored 88.2% on SEC-bench Pro, which tests vulnerability discovery; 54.8% on ExploitGym, which tests exploitation; and 71.0% on CWE-bench, which tests software patching.
Google conducted those cybersecurity evaluations without guardrails. On internal tests, Google says Carbon was more than twice as successful as Argon at generating secure code.
The company says its teams have seen Carbon work for days on multistep engineering problems, including identifying and fixing subtle bugs across large codebases. Google also reports fewer unsupported claims and fewer high-severity misalignment alerts in its internal tests than with Argon. Carbon scored 59 on the updated Artificial Analysis Intelligence Index, placing first amongst all models.

Google plans to expand access gradually through its Fairwind Program before making Carbon available to developers, businesses and consumers.
The introductory API price will be $2 per million input tokens and $10 per million output tokens, rising to $4 and $20, respectively, after the introductory period. Google has not given a date for wider availability.
TheWhiteBox’s takeaway:
Pretty strong showing at a bad moment for the company after the failed earnings (see above markets section).
My problem with Google, and I’m saying this as a shareholder, is that these releases are ‘non-releases’, as almost no one has actually tested Gemini 4 models yet. They are announced, but not released.
The caveat is that the model seems to be very expensive to run. Yes, it sets a new record on the artificial analysis benchmark, even beating Claude, but the price tag is not cheap, which is not great considering Google was long known for being the most cost-effective provider. Sadly, that crown has long been lost to OpenAI.

Closing Thoughts
This week has been dominated by market data, with Anthropic’s data coinciding with earnings season. Results are good, but markets haven’t shaken off the fears that AI is running ahead of its skis.
The fact that the FED chose not to raise interest rates will prevent yields from rising further, which should help the trade. But this industry is getting a reality check it's long overdue for.
AI can prove any conjecture it wants. The ETH discovery is amazingly cool, no doubt. But formally proven conjectures don’t pay the bills. This industry needs killer applications that stop prices from dropping and increase conversions. Because today, the vast, vast majority of society doesn’t pay a single dollar for its AI services.
In the process, we keep spending out of our minds, expecting people to convert. But I feel like I’ve been saying this not just for too many months, but too many years.
Markets are growing tired of waiting, and the macroeconomic context is that most countries are massively polarized politically, and government debt surging across the board isn’t something that increases AI’s chances.
And sadly, Anthropic’s S1 hasn’t reversed course on any of these doubts. And if Anthropic couldn’t, what will?
Until the next one!

Give a Rating to Today's Newsletter
For business inquiries, reach me out at [email protected]

