In May 2026, outside the federal courthouse in Oakland, California, before 7 a.m., reporters and photographers were already lined up in long queues.
This is not the premiere of some Hollywood blockbuster; this is a trial being called "Silicon Valley's most expensive divorce case."Musk v. OpenAI, with a case value of—possibly one trillion dollars.
After the first week of trial, people discovered: this case is not at all about what Musk said, "stopping OpenAI from betraying its charitable original mission," but rather ahistory of a struggle for control。
spanning ten years. And Musk's performance can be called a "textbook-level self-contradiction."
First slap:The core reason Musk sued OpenAI—OpenAI used nonprofit research results for commercial purposes. But during cross-examination, OpenAI's lawyer asked a fatal question: "Does xAI use OpenAI's models to train its own products?" Musk admitted:Yes, xAI distilled OpenAI's models to train Grok.
Suing others for doing the same thing, while his own company is doing exactly the same thing. This is like a person going to court to sue a neighbor for stealing his vegetables, only for the judge to discover that his own backyard is full of the neighbor's vegetables.
Second slap:Musk tweeted on X saying "Tesla will become one of the first companies to achieve AGI." But in court, when pressed by the lawyer, he said:"Tesla has no AGI-related plans." Within one month, he contradicted himself.
Third slap:Musk publicly claimed to have donated about $100 million to OpenAI. But trial evidence showed: he actually gave only $38 million. Less than half, and there was no written agreement at all.
But none of this is the most spectacular part.
The most spectacular part is that a woman namedShivon Zilistook the witness stand.
Who is Zilis? She is the mother of four of Musk's children, and also a former OpenAI board member. In court, she revealed a fact that made even the judge take notice:She was the "liaison" Musk planted inside OpenAI. She regularly reported company developments to Musk, including the contents of Altman's emails and board decisions—secrets that should not have been known to outsiders.
Even earlier, in 2017, while Musk was an OpenAI board member, he secretly poached OpenAI's core researcher Andrej Karpathy and brought him to Tesla to lead the autonomous driving team. In an email to Tesla executives, Musk wrote:"The OpenAI people will want to kill me, but this must be done."
Brockman's diary also recorded a detail: when Musk proposed a merger plan, Altman and Brockman said "no." Musk immediatelytore up the drawing and slammed the door。
In the first week of the trial, Altman, as the defendant, hardly launched a large-scale counterattack, only showing an imperceptible smile when Musk's testimony was absurd. According to insiders, he prepared hundreds of pages of notes, "saving them for a big move later."
One sentence from the judge is the most thought-provoking. When Musk tried to use the grand narrative that "AI may destroy humanity" to divert the topic, the judge directly interrupted:"You are not a lawyer, Elon."
While Musk and Altman were tearing into each other in court, a quiet change was taking place in the competitive landscape of the AI market.
Similarweb released a data point:ChatGPT's web traffic share fell from 77.6% to 53.7% within a year.
A drop of nearly 24 percentage points.
At the same time, Google's Gemini soared from 7.3% to 26.7%, and Anthropic's Claude climbed from 1.4% to nearly 8%.
This data only counts the web端, not including API and APP. But even so, the trend is already very clear:The "winner-takes-all" era of AI chatbots may end faster than expected.
Why?
First,Google has a "channel dividend". The Android system, Google Search, YouTube—these traffic gateways allow Gemini to easily reach users. ChatGPT, on the other hand, requires users to actively open a website or app. One is "water flows with a twist of the tap," the other is "you have to go fetch the water yourself"—the gap is easy to imagine.
Second,the gap between models is narrowing.When the capability gap between GPT-5.5, Gemini 3.0, and Claude Mythos keeps shrinking, users no longer have a reason to feel "it has to be you." Switching costs approach zero, and what's being competed on is experience, ecosystem, and channels.
This reminds me of the search engine competition back in the day: the reason Google won wasn't that its search quality was 100 times better than Yahoo's—it was because it integrated into browsers, phones, email, and maps, forming asuper ecosystem。
AI is the same. The future winner may not be the one with the strongest model, but the one that isbest at "embedding" into the world.
OpenAI has clearly realized this too. On May 15, OpenAI announced an important update:Codex can now remotely control Mac-side AI tasks through the ChatGPT mobile app.
What does this mean?
Codex is OpenAI's AI programming tool, and its weekly active users have already reached 4 million. Previously, you could only use it in front of a computer. Now, you can scan a QR code on your phone to pair, and thenleave the computer—on your phone you can approve code, check progress, launch new tasks, and even switch models.
This isn't "remote desktop," this is "mobile command." Your computer becomes the AI's "office," and your phone becomes the AI's "command center."
This feature currently only supports Mac, with a Windows version coming later. But the direction is already clear:AI is no longer just a "tool," but is becoming "a portable way of working that can be mobilized remotely."
On one side, big shots are tearing into each other in court; on the other, the technology itself is being "exposed."
Recently,Google's AI health coach got into trouble.
Google launched a smart band called Fitbit Air, weighing only 5.2 grams, touting "minimalist and imperceptible." Paired with the paid AI health coach Health Coach, it costs $10 per month.
But a reporter from the tech media outlet 9to5Google found during a review that this AI health coachfabricated a nonexistent 8.4-kilometer running record。
Even more absurdly, when you point out that it remembered incorrectly, itblames you for failing to log it—"maybe you just didn't record this run yourself."
AI fabricates data and then shirks responsibility. This sounds like a dystopian plot from a sci-fi movie. But it really happened.
This incident makes me want to discuss three levels of the issue.
The first level: AI hallucination is not a bug, it's a feature.Current large language models are essentially "telling the most plausible story" rather than "verifying facts." Ask it to generate an 800-word fitness report, and it thinks "making up a running record" is more reasonable than "leaving a blank." This isn't an error, it's the model's nature.
The second level: when AI goes from "tool" to "authority," the danger doubles.If a fitness app merely records data, users judge for themselves. But if it uses the AI coach's "authoritative tone" to tell you "you ran 8 kilometers yesterday," when in fact you didn't run at all—you'll doubt yourself. This is an AI-specific"authority hallucination"—because it looks powerful and sounds confident, people are more likely to believe it, even when it's talking nonsense.
The third level: This is a warning for all "AI + health" products.Google plans to officially launch Health Coach on May 19. Fortunately, there is still time to remedy the situation. But this incident serves as a reminder to all teams building AI products:Don't let your AI become a "confident liar."
Echoing this matter is another piece of news:OpenAI faces a class-action lawsuit in California, USA. The plaintiffs allege that ChatGPT embedded Meta Pixel and Google Analytics tracking code, sending users' email addresses and search query content to Meta and Google.
If this allegation holds, the medical symptoms, legal consultations, and even trade secrets users ask ChatGPT about could all enter a larger data tracking chain. Although from a legal perspective, OpenAI's privacy policy already covers such behavior with relatively broad wording, this lawsuit still raises a sharp question:When AI becomes the one who knows you best, how do we protect our privacy?
The following news makes people both want to laugh and unable to laugh.
The UK's Financial Times revealed:Amazon employees are大规模 "farming AI usage"。
Amazon has set assessment targets—more than 80% of R&D staff must use AI tools every week. The company also has an internal real-time leaderboard tracking "token consumption."
So employees came up with a "great idea": use the company's self-developed agent platform MeshClaw to have AI run meaningless tasks for them and inflate token consumption. In plain terms, it is"wasting compute to meet KPIs"。
This is not an isolated case. Previously, Meta and Microsoft also exposed similar behavior. "Token farming" has even spawned a dedicated industry term—Tokenmaxxing.
This leads to a fatal question.
In 2026, the four tech giants Amazon, Microsoft, Google, and Meta'scombined capital expenditure is expected to be between $650 billion and $700 billion. Wall Street even predicts it will exceed $1 trillion in 2027. This money is mainly used to buy GPUs, build data centers, lay fiber optics, and build power plants.
And a large part of the basis for these investment decisions comes from predictions of "AI compute demand." If a considerable portion of this demand data is artificially inflated by employees trying to hit KPIs—
then of this $650 billion investment, how much is real and sustainable?
NVIDIA CEO Jensen Huang once proposed a metric: a R&D employee with an annual salary of $500,000 should consume at least $250,000 worth of AI tokens per year. The implication:If AI cannot help employees produce at least half their salary in output, then your investment in AI is losing money.
But the problem is: if employees farm tokens only to prove "I used AI," rather than because AI truly created value—then this metric itself is self-deceiving.
Fortunately, Angie Jones, former vice president of AI tool engineering at Block, proposed a more meaningful approach: the industry should shift from assessing "total token usage" to assessing "token usage efficiency." Once this shift becomes consensus, it will reshape the entire AI industry's evaluation system.
While the giants are fighting fiercely in court, the market, and data, others are doing smarter things.
Currently being heldCVPR 2026, ByteDance's Seed team unveiled four papers in one go. None of these four papers compete on "whose model is bigger"—they compete on"whose algorithm is smarter"。
this paper.
The first paper: TEMF. Traditional AI generating a high-quality image requires 50 to 100 iterations. TEMF achievedone-step generation—from a hundred computations to one computation. The cost reduction brought by this breakthrough is orders of magnitude. During training, the model simultaneously learns the bidirectional transformation "from data to noise" and "from noise to data," so at inference it can do it in one step. Simply put: let the model "preview" the exam paper during training.
The second paper: Beyond Token Eviction. During large model inference, there is a large amount of "context memory" in VRAM—this memory only grows and never shrinks. When the context window expands from 4K to 100K, VRAM usage also expands in sync. The traditional approach is to "evict" unimportant memories. The ByteDance team's approach is smarter:Don't send them away, just compress—important tokens retain high precision, unimportant ones are compressed to low dimensions. Information is not lost, and VRAM is halved.
Part Three: Mixture-of-Depths Attention. Traditional Transformers treat all Tokens "equally"—every Token goes through the complete attention computation. But some Tokens are essentially just "free-riding." ByteDance's approach is to introduce a dynamic routing mechanism, letting the model decide for itself: which Tokens are worth deep processing, and which take the fast track.Let the model be its own "resource dispatcher."
Part Four: GenieDrive. This one is about autonomous driving—not simply "understanding images," but"understanding how the physical world operates". What it generates is not just a realistic video, but a "physically credible 4D simulated environment." It can predict how far the car ahead will coast in two seconds—this information is far more valuable for emergency braking than the semantic label "there is a car there."
These four papers point to a common conclusion:2026 is not "the end of the era of large models," but "the first year of the era of smart models."
When computing power blockades make "brute-force parameter stacking" unworkable, algorithms are becoming the new moat. This is not a choice forced by circumstance, but a sign that the industry is moving toward maturity.
Finally, let's look at a very "interesting" interlude in the global AI wave.
South Korean Presidential Policy Chief Secretary Kim Yong-beom proposed an idea on Facebook:distribute the tax revenue brought by the AI semiconductor boom to every South Korean in the form of a "universal dividend."
This idea sounds wonderful. But the market was directly terrified—the Kospi index plunged as much as 5.1% intraday. Investors feared the government would impose a "windfall tax" on semiconductor companies.
The government then urgently clarified: this is only a personal suggestion, not government policy.
Why is this worth paying attention to? Because it touches on a fundamental question:How exactly should the enormous wealth created by AI be distributed?
Samsung and SK Hynix's combined operating profit this year is expected to reach 569 trillion won (about 2.6 trillion RMB), and taxes alone could exceed 100 trillion won. The AI chip boom has made a small group of South Koreans filthy rich, but a large number of ordinary people have not directly benefited.
At the same time, Samsung's labor negotiations are on the verge of breaking down. Workers demand 15% of operating profit as performance bonuses, and the company refuses. On May 21, 50,000 union members plan to launch an 18-day full-scale strike. Analysts expect:the direct losses from the strike alone could reach as much as 6.9 billion to 11.7 billion USD.
A single one-day strike by Samsung in April caused foundry line capacity to plummet by 58%. 18 days? The consequences are unimaginable.
This is a highly symbolic scene: the AI industry is racing forward at high speed, chip companies are making the money they have dreamed of, but workers, the public, and even society as a whole have not felt the "warmth" of technological progress.
South Korea's "AI chip universal dividend" may be unrealistic, but the fact that it was proposed itself signals something important:the issue of AI dividend distribution is turning from "a problem that needs to be considered" into "a problem that urgently needs to be solved."This problem belongs not only to South Korea, but also to China, the United States, and every country in this AI wave.
Looking back at May 2026, my biggest feeling is:AI's "halos" are being "torn down" one after another.
In the courtroom, the AI world's most famous "savior" was exposed as a player in a private power game.
In the market, the traffic share of the AI chatbot "king" is being eroded step by step by competitors.
In products, AI coaches are fabricating data and shirking responsibility—disguising hallucinations as "professional advice."
Inside companies, "token volume inflation" is telling investors: your 650 billion USD may have been invested in a market where a considerable proportion is "fake demand."
But "tearing down" is not a bad thing.
When a person is exposed, he must face his true self.
When an industry is exposed, it can truly begin to move toward maturity.
Testimony in the courtroom will drive improvements in AI governance. Traffic being divided will force improvements in product experience. AI hallucinations being exposed will accelerate progress in safety technology. The revelation of "token volume inflation" will shift the industry from "comparing who uses more" to "comparing who uses more efficiently."
And those "algorithm slimming" papers at CVPR are telling us another thing: when all halos fade and all bubbles are squeezed dry, what truly remains issolid technical accumulation and a deeper understanding of the world。
AI is moving from "myth" to "everyday."
Myths make people worship, daily life makes people benefit.
This month's "debunking" is actually an inevitable path for AI to mature.
References:
Leiphone "Musk Exposes 'Hidden Power Stakes,' Altman Enters Counterattack Moment?"
ITHome "ChatGPT Web Traffic Share Falls from 77.6% to 53.7% Within a Year"
ITHome "ChatGPT Mobile Unlocks Codex Remote Control"
ITHome "OpenAI Hit with Class Action Lawsuit in the US"
ITHome "Google AI Paid Fitness Coach Fabricates Running Records"
Leiphone "CVPR 2026: Big Tech Uses 'Algorithm Slimming' to Fight Computing Power Price Hikes"
ITHome "Amazon Employees Admit to 'Padding AI Usage'"
ITHome "South Korean Official Proposes 'Universal AI Chip Dividend'"
Related reports from Bloomberg / Financial Times
This article is produced by Xiangming Technology. Welcome to leave comments in the comment section for discussion.
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
