Techin' Geek

The newest update for NotebookLM introduces Deep Research and compatibility with Sheets, Word, PDFs, and images, transforming Google’s AI notebook into a more comprehensive study center where your documents, references, and summaries coexist.

Following a year filled with aggressive negotiation and speculation surrounding a potential IPO, the financial examination of OpenAI is becoming more rigorous. Confidential documents acquired by tech commentator Ed Zitron shed more light on OpenAI’s finances — particularly its earnings and computational expenses over recent years.  

Zitron disclosed this week that in 2024, Microsoft secured $493.8 million from OpenAI in revenue share payments. That figure soared to $865.8 million in the first three quarters of 2025, per the documents he reviewed.

OpenAI reportedly provides Microsoft with 20% of its revenue due to a prior agreement in which the software titan invested more than $13 billion in the influential AI startup. (Neither the startup nor individuals in Redmond have publicly validated this percentage.)

However, this situation gets somewhat complicated, as Microsoft also distributes revenue to OpenAI, contributing approximately 20% of the revenues from Bing and the Azure OpenAI Service, a source familiar with the situation informed TechCrunch. Bing operates with OpenAI’s technology, while the OpenAI Service markets cloud access to OpenAI’s models to developers and enterprises.  

The source also indicated to TechCrunch that the leaked payments pertain to Microsoft’s net revenue share, not the gross revenue share. In simpler terms, they don’t encompass what Microsoft paid to OpenAI through Bing and Azure OpenAI royalties. According to this source, Microsoft deducts those amounts from its internally reported revenue share figures.

Microsoft doesn’t specify how much it earns from Bing and Azure OpenAI in its financial reports, making it challenging to estimate how much the tech behemoth is returning.

Nonetheless, the leaked documents offer insights into the most sought-after company in the private market currently — revealing not only its revenue but also its expenditures in light of that revenue.  

Thus, using the widely circulated 20% revenue-share figure, it can be inferred that OpenAI’s revenue was at least $2.5 billion in 2024 and $4.33 billion during the first three quarters of 2025 — though it is likely higher. Previous reports from The Information estimated OpenAI’s 2024 revenue at about $4 billion, and its revenue in the first half of 2025 at $4.3 billion.  

Altman recently stated that OpenAI’s revenue exceeds reports of $13 billion annually, predicting it will surpass $20 billion in annualized revenue run rate by year-end (a projection, not guidance on actual revenue), with a possibility of reaching $100 billion by 2027. 

According to Zitron’s examination, OpenAI may have incurred approximately $3.8 billion in inference expenses in 2024. That expense grew to around $8.65 billion in the first nine months of 2025. Inference refers to the computing power utilized to operate a trained AI model for generating responses.  

Historically, OpenAI has predominantly depended on Microsoft Azure for compute access, although it has also established agreements with CoreWeave and Oracle, and more recently, with AWS and Google Cloud. 

Prior reports estimated OpenAI’s total compute expenditure at roughly $5.6 billion for 2024 and its “cost of revenue” at $2.5 billion for the initial half of 2025.  

A source knowledgeable about the matter told TechCrunch that while OpenAI’s training expenses are mostly non-cash — indicating they are covered by credits Microsoft granted OpenAI as part of its investment — the company’s inference expenses are largely cash-based. (Training pertains to the computing resources needed for the initial training of a model.)

Although this does not present a complete picture, these figures suggest that OpenAI might be spending more on inference costs than it brings in through revenue. 

These implications are likely to fuel the ongoing discussions about the AI bubble that have permeated conversations from New York City to Silicon Valley. If the model leader OpenAI continues to operate at a loss while running its models, what could this signify for the significant investments at astonishing valuations in the broader AI sector?

OpenAI declined to provide comments. Microsoft did not respond to TechCrunch’s inquiry for comment.

Do you have a confidential tip or documents? We’re reporting on the inner workings of the AI industry — from the companies shaping its future to those affected by their decisions. Reach out to Rebecca Bellan at [email protected] or Russell Brandom at [email protected]. For secure communication, you can contact them via Signal at @rebeccabellan.491 and russellbrandom.49.

Andy Konwinski voices alarm over the waning U.S. leadership in AI research to China, deeming this transition an “existential” threat to democracy. Konwinski co-founded Databricks and the AI research and venture capital firm Laude.

“If you engage with PhD students at Berkeley and Stanford focusing on AI presently, they’ll mention having encountered twice as many intriguing AI concepts in the past year from Chinese enterprises compared to American ones,” Konwinski remarked during his appearance at the Cerebral Valley AI Summit this week.  

Alongside his investments via Laude, the venture fund he established last year with NEA alumnus Pete Sonsini and Antimatter CEO Andrew Krioukov, Konwinski also leads the Laude Institute, an accelerator providing grants to academics.

Prominent AI laboratories such as OpenAI, Meta, and Anthropic continue to make significant advancements, yet their contributions largely remain proprietary rather than open source. Furthermore, these corporations are attracting top academic talent by offering salaries in the millions that exceed what these professionals can earn in academia.

Konwinski contended that for ideas to genuinely thrive, they must be freely shared and debated within the broader academic community. He emphasized that generative AI arose directly from the Transformer architecture, a crucial training technique introduced in a openly accessible research publication.

“The first country to achieve the next breakthrough at the ‘Transformer architectural level’ will possess the upper hand,” Konwinski asserted.

Konwinski points out that in China, government support and encouragement for AI innovation, whether from labs like DeepSeek or Alibaba’s Qwen, promotes open sourcing, permitting others to build on these innovations, which, he argues, will inevitably result in further breakthroughs.  

He believes this contrasts sharply with the U.S., where, as he describes, “the sharing of knowledge among scientists that we have always experienced in the United States has diminished.”

Konwinski suggests that this phenomenon not only threatens democracy but also poses a business risk to key U.S. AI laboratories. “We’re consuming our own future resources; the source is drying up. Fast-forward five years, the major labs will also be adversely affected,” he stated. “We need to ensure that the United States remains at the forefront and open.” 

Tesla has provided the most thorough insight into the operation and comparative safety of its sophisticated driver-assistance system, shortly after Waymo’s co-CEO Tekedra Mawakana urged companies at TechCrunch Disrupt to disclose more information.

On a newly added section of its website, Tesla asserts that in North America, users of the company’s Full Self-Driving (Supervised) software are logging approximately 5 million miles between serious collisions and about 1.5 million miles before a minor crash.

This figure is significantly below the national average as per the data provided by the National Highway Traffic Safety Administration (NHTSA). According to this data, individuals encounter a major accident every 699,000 miles, and a minor one every 229,000, at least per Tesla’s interpretation.

Tesla has been publishing “vehicle safety reports” on a quarterly basis for some time. However, these reports have faced criticism for being inadequate. Furthermore, Tesla has disclosed nearly no information regarding the safety performance of the Robotaxi pilot it has been conducting in Austin, Texas, this year, where employees continue to occupy the driver’s seat for safety oversight.

Waymo, the top robotaxi firm in the U.S. currently based on the number of cars in service and clients served, has shared comprehensive data indicating that its vehicles are around 5 times safer than human operators, and 12 times safer regarding pedestrian interactions. At last month’s Disrupt conference, Mawakana was posed the question of other companies she believed were contributing to road safety.

“I can’t identify those on the list since they aren’t revealing information about their fleets,” mentioned Mawakana, while not specifying Tesla.

“I believe there is an obligation, if you are going to operate vehicles on public roads, and if you are going to eliminate the driver from the vehicle, and have someone in a separate location monitoring the fleet who can take control of these vehicles, you must be transparent about what’s occurring,” she added. “If you’re not being transparent, then in my opinion, you’re not fulfilling your responsibility to truly make the roads safer.”

Waymo did not respond promptly to a comment request on Friday regarding Mawakana’s view on whether Tesla’s newly released data suffices.

A recurrent critique of Tesla’s quarterly safety reports is their emphasis on Autopilot, which is a less sophisticated driver-assistance system than the Full Self Driving (Supervised) software, or FSD — which, despite its name, does not render a vehicle fully autonomous. Autopilot was designed for highway use, where the crash rate is usually lower (including minor collisions).

Tesla has at last detailed all this information. The fresh segment on Tesla’s website reports that drivers utilizing FSD cover about 2.9 million miles before major collisions, while NHTSA data states that all drivers cover around 505,000 miles per major crash. Tesla claims FSD users travel approximately 986,000 miles between minor collisions, whereas NHTSA data depicts that all drivers cover about 178,000 miles per minor accident.

Tesla is now also clarifying how it interprets these terms for the first time.

The automaker is referencing the Federal Motor Vehicle Safety Standards, specifically 49 C.F.R. § 563.5. Tesla characterizes “major collisions” as incidents with more severe impacts where a vehicle’s airbags “or other non-reversible pyrotechnic restraints” have been deployed. The company indicates that if FSD was active “at any moment within five seconds leading up to a collision event,” that crash is included in this data set.

“This calculation guarantees that our reported collision rates for FSD (Supervised) reflect not only collisions occurring while the system is actively managing the vehicle but also cases where a driver might disengage the system or situations where the system aborts autonomously shortly before impact,” Tesla states.

In its FAQ section, Tesla mentions that it will refresh the data quarterly and that it will “illustrate a rolling twelve-month aggregation of miles and collisions to stay relevant to current trends and advancements.” The company states it will not share additional information, such as injury statistics, because it is automatically collecting this data from its vehicles.

“Instead, Tesla concentrates on objective and programmatic metrics such as collision frequency and airbag deployment rates. Airbag activations serve as a dependable indicator for collision severity,” the company explains.

To be frank, the majority of tech wearables tend to be cumbersome and not what one would consider attractive. However, if you’re frequently adorning something on your fingers, you likely want it to be visually appealing and possibly integrate seamlessly with your other jewelry. This is the concept behind Oura’s latest smart ring, the Oura Ring 4 Ceramic.

The Ring 4 Ceramic represents Oura’s inaugural series of smart rings that does not include metallic finishes. Instead, they are constructed from zirconia ceramic, a material that is more resistant to wear, which Oura claims provides both elegance and comfort.  

I’ve been using the Oura Ring 4 Ceramic over the last three weeks, and here’s what I’ve discovered.  

Before getting into the details, it’s important to point out that as the software in the Oura Ring 4 Ceramic mirrors that of the Oura Ring 4, I won’t be exploring the software and hardware aspects extensively. For that, you may refer to our Oura Ring 4 review.  

Appearance and sensation 

The Oura Ring 4 Ceramic is available in four shades: Midnight (dark blue), Cloud (white), Tide (light teal), and Petal (light pink). The price of the ring is $150 higher than the base price of the standard Ring 4, priced at $500, and it comes in sizes 4-15.  

I selected Midnight. Although the hue is classified as dark blue, it may look black from a distance — or even a dark green, according to some. At first glance, the ring doesn’t resemble a tech wearable, which is a plus if you prioritize looks without compromising on functionality.

Compared to the original Ring 4, the Ceramic variant is slightly more robust, but the distinction is barely perceptible. The titanium variant has a thickness of 2.88 mm, while the ceramic variant stands at 3.51 mm.

Regarding weight, the ring can range from 5.1 to 8.1 grams, based on the size you select, making it heavier than the Titanium Ring 4, which weighs between 3.3 and 5.2 grams.

Even though the Oura Ring 4 Ceramic is thicker and heavier than its predecessor, I found it to be both comfy and lightweight, particularly when compared to a conventional metal band. It feels much like a regular ring, just a bit thicker.

The ring has a smooth texture and does not collect fingerprints as readily as the titanium options from the Ring 4. Due to the fact that zirconia ceramic is warmer and less slippery than titanium, the Ring 4 Ceramic may feel somewhat “sticky” at times, resulting in the outer surface occasionally adhering to your fingers, but I got used to the sensation pretty fast.  

Oura mentions that the ring’s color results from natural minerals within the ceramic itself, ensuring it remains vibrant and does not fade over time. (Of course, since I’ve only had the ring for three weeks, I can’t attest to its long-term durability, but thus far, the color is consistent with what it was on day one.)

Durability 

Oura cautions that the Ring 4 Ceramic is susceptible to scratches from softer metals, such as those found in cookware or weights. The company states that these scratches can be managed with the Polishing Pad provided with the ring. (However, the review unit I received from Oura did not come with the Polishing Pad.)

The company further suggests wearing the ring on your non-dominant hand to minimize impacts. Nevertheless, the ring felt just right on my dominant index finger, so that’s where I wore it for the past three weeks. This resulted in increased potential exposure to damage in a shorter time frame than a standard user might experience.

On the first day I wore the ring, I put it through a rigorous travel day at TechCrunch Disrupt, our yearly conference in San Francisco. I was unsure how well the ring would endure as I managed my luggage, inadvertently knocking the ring against several counters (I’m rather clumsy!), or allowing it to graze the airplane window when I leaned in to rest.  

As it turned out, the ring emerged from all that without a scratch. There were multiple instances when I thought for sure it had been marked — like when I grasped a rusty railing while ascending the stairs of the Point Arena Lighthouse, forgetting I wore the ring, or neglecting to take it off before washing dishes. However, the ring remained free of any visible marks or lasting damage.

It’s important to note that while the ring is water-resistant up to 100 meters, making it suitable for activities like washing dishes or showering, I chose to remove it during dishwashing because Oura advises that it could get scuffed when in contact with softer metals.

Given that I’ve had the ring for less than a month, I wouldn’t be surprised if it picks up some scuffs or dents over time, especially because of its placement on my dominant hand.

Concluding thoughts  

Aside from its appearance, the ring proved to be quite beneficial in providing insights into my heart health, sleep patterns, and stress levels through the Oura app. Each morning, I found myself eager to check my “readiness score,” calculated using both short-term and long-term metrics to indicate how geared up you are for the upcoming day.

Moreover, the battery life is performing well, requiring only weekly charging, which aligns with Oura’s claim of a battery lasting between five and eight days.

Deciding between a titanium or ceramic ring ultimately hinges on your design preferences and budget considerations. With color and material as the sole distinctions between the two, it boils down to whether you prefer a sleek, contemporary finish or the typical metallic smart ring aesthetic. 

If you’re after that splash of color and style, then the Oura Ring 4 Ceramic could be the perfect smart ring for you.  

OpenAI has announced that ChatGPT will now refrain from using em dashes if directed to do so. This distinctive feature, which has come to symbolize AI-generated text, has been seen widely in recent months across school papers, emails, comments, customer service interactions, LinkedIn posts, online discussions, advertising content, and more. The presence of the em dash has prompted some to criticize those authors for relying lazily on an AI chatbot for assistance.

Of course, many advocates of the em dash argue that it has been part of their writing style long before large language models adopted this punctuation mark. Nonetheless, chatbots’ frequent misuse has rendered the so-called “ChatGPT hyphen” an unwelcome characteristic in any writing, even though it is not a conclusive indicator of text crafted by generative AI.

This issue had perplexed OpenAI for a while, as users of ChatGPT found it challenging to stop the chatbot from employing the symbol, even when they explicitly requested it not to.

Now, OpenAI’s CEO Sam Altman states that the issue has been resolved. In a message posted on X, Altman noted, “If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it’s supposed to do,” describing the update as a “small-but-happy win.”

Small-but-happy win:

If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it’s supposed to do!

— Sam Altman (@sama) November 14, 2025

The company elaborates in a message on Threads (where it compelled ChatGPT to apologize for “ruining the em dash”) that ChatGPT will be more adept at avoiding the em dash if you instruct it not to through the custom instructions in your personalization settings. While it may not automatically eliminate the em dash from its responses, you will have increased control over how often it appears.

Five individuals have admitted guilt for assisting North Koreans in defrauding U.S. firms by impersonating remote IT personnel, as stated by the U.S. Department of Justice (DOJ) on Friday.

The five individuals are charged with acting as “facilitators,” enabling North Koreans to secure employment by using their genuine identities, or the false and stolen identities of over a dozen U.S. citizens. The facilitators also accommodated company-issued laptops in their residences throughout the U.S. to create the impression that the North Korean employees were residing locally, as per the DOJ’s news release.

These activities impacted 136 U.S. companies, yielding $2.2 million in revenue for Kim Jong Un’s regime, according to the DOJ.

This latest series of guilty pleas is part of an ongoing initiative by U.S. authorities aimed at undermining North Korea’s capacity to profit from cybercrime. For years, North Korea has effectively infiltrated numerous Western enterprises as remote IT personnel — as well as investors and recruiters — to finance its internationally condemned nuclear weapons program. In recent years, the U.S. government has retaliated by indicting those involved in the scheme and imposing sanctions on global fraud networks. 

“These prosecutions clearly indicate that the United States will not allow [North Korea] to fund its weapons initiatives by exploiting American businesses and workers,” stated U.S. Attorney Jason A. Reding Quiñones in a news release. “We will continue collaborating with our partners throughout the Justice Department to expose these frauds, recover embezzled funds, and pursue every individual who supports North Korea’s operations.”

Three of the individuals — U.S. citizens Audricus Phagnasay, Jason Salazar, and Alexander Paul Travis — each entered a guilty plea to a single charge of wire fraud conspiracy. 

Prosecutors claim that the three assisted North Koreans posing as legitimate IT personnel, whom they recognized as working outside of the U.S., to utilize their identities to gain employment, helped them remotely access company-issued laptops located in their homes, and assisted the North Koreans in passing vetting processes, such as drug screenings. 

Travis, identified by prosecutors as an active member of the U.S. Army at the time of the scheme, received over $50,000 for his activities, while Phagnasay and Salazar earned at least $3,500 and $4,500, respectively. The scheme resulted in U.S. companies disbursing approximately $1.28 million in salaries, most of which was sent to the North Korean IT workers abroad, according to the DOJ.

The fourth U.S. citizen who admitted guilt is Erick Ntekereze Prince, who operated a firm named Taggcar, which falsely provided U.S. companies with “certified” IT workers whom he knew were operating outside the country and utilizing stolen or fake identities. Prince also maintained laptops with remote access software at various locations in Florida, earning over $89,000 for his services, as stated by the DOJ. 

Another participant in the scheme who pleaded guilty to one count of wire fraud conspiracy and an additional count of aggravated identity theft is Ukrainian national Oleksandr Didenko, whom prosecutors accuse of appropriating U.S. citizens’ identities and selling them to North Koreans to facilitate employment at over 40 U.S. companies. 

According to the press release, Didenko earned several hundred thousand dollars for this service. Didenko consented to forfeit $1.4 million as part of his plea agreement.

The DOJ also revealed that it has frozen and seized more than $15 million in cryptocurrency that was stolen in 2023 by North Korean hackers from multiple crypto platforms. 

Crypto firms, exchanges, and blockchain enterprises have become prime targets for North Korean hackers, who pilfered over $650 million in crypto in 2024 and more than $2 billion up to this year.

Boeing has reached an agreement with the startup Charm Industrial to extract 100,000 metric tons of carbon from the air.

Charm gathers waste from agriculture and forestry and applies heat to convert it into a substance they refer to as “bio-oil,” a complex mixture of hydrocarbons that it injects underground, including into decommissioned oil wells. Once captured, Charm can market carbon removal credits to businesses. Axios was the first to report on the agreement between the startup and Boeing.

The aviation industry has made minimal strides in reducing its carbon emissions. This has prompted companies in the field to look for alternatives. Carbon removal has surfaced as a viable option since it could be cheaper than shifting to sustainable aviation fuels.

A study indicated that by 2050, the aviation sector will need to invest at least $60 billion in carbon offsets to achieve net zero emissions.

Charm is also capable of generating biochar, a material that, when applied to agricultural fields, can enhance soil fertility, although those initiatives are still in the early stages, based on information from the carbon removal registry Isometric.

The financial aspects of the deal were not made public. Two years prior, Charm sold 112,000 carbon removal credits to Frontier, the advanced market commitment, for $53 million, equating to around $470 per metric ton. Charm has expressed its goal to lower the price to approximately $50 per metric ton.