Financial industry leaders convened to discuss the cyber risks posed by Anthropic’s latest AI model after it reportedly found weaknesses in every major computer operating system. That’s a sentence that would have read as science fiction five years ago. It’s now a compliance meeting.

The specifics of what was found, and how, remain unclear from public reporting — which is its own kind of signal. When that kind of information circulates first in closed industry sessions rather than public disclosures, it suggests the vulnerabilities are either still being patched, or the exposure is broad enough that nobody wants to start a countdown clock before fixes are in place. Either way, the episode is a clean illustration of the dual-use problem at the core of frontier AI: the same capability that finds vulnerabilities defensively is also the one that finds them offensively. The institutions meeting about this risk are right to take it seriously. Whether they’re moving fast enough is a different question.

Microsoft, Google, Amazon and other major tech companies are locking in contracts with nuclear startups to secure reliable power for AI data centers. The rationale is straightforward: AI infrastructure runs continuously, consumes enormous amounts of electricity, and needs stable baseload power that solar and wind can’t reliably provide without storage that doesn’t yet exist at sufficient scale. Nuclear, which had been commercially stagnant for decades, is suddenly the destination of serious capital.

Andy Jassy disclosed in his annual shareholder letter that Amazon’s AI services at AWS are running at a $15 billion annual revenue rate. He also revealed that the company’s custom chip business — Graviton and Trainium — has crossed a $20 billion annualized run rate, roughly double what the company cited earlier this year.

Amazon stock jumped more than 5% on the news. The S&P closed at 6,824 and the Nasdaq at 22,822 on Thursday as the letter gave investors a concrete number to attach to years of capital expenditure narrative.

Atlassian announced it is laying off approximately 10% of its global workforce — around 1,600 people — to redirect resources toward AI development and enterprise sales. The company estimates restructuring costs up to $236 million. It simultaneously replaced its CTO with two new AI-focused executives in the role.

This is now a recurring pattern in enterprise software. A company that built its reputation on collaborative tools announces it is dismantling some of its human workforce to build AI replacements for that workforce’s output. The framing is always “investment in the future.” The experience for the people being let go is something else.

Meta has signed a new deal to spend an additional $21 billion with CoreWeave between 2027 and 2032, on top of a prior $14.2 billion commitment. Total exposure: $35.2 billion to a single GPU infrastructure provider. Meta’s projected capital expenditures for 2026 alone sit at $115 to $135 billion — nearly double 2025 levels.

The logic is portfolio hedging. Meta is building its own facilities (including a major Texas data center) while simultaneously contracting with CoreWeave for scalable capacity it can access immediately without the 18-month construction timeline. CoreWeave benefits by diversifying away from Microsoft, which previously represented a dominant share of its revenue.

Microsoft announced a $10 billion investment in Japan spanning 2026 through 2029 — AI infrastructure, cybersecurity partnerships, and a commitment to train over one million engineers and workers by 2030.

It’s the follow-on to a $2.9 billion investment in April 2024. The new package is organized around three pillars: Technology, Trust, and Talent. Microsoft is also joining Japan’s Kyushu Semiconductor Human Resource Development Consortium — the first major international tech company to do so.

OpenAI has passed $25 billion in annualized revenue and is reportedly taking early steps toward a public listing, potentially as soon as late 2026. The company is currently valued at $852 billion — not a typo. Rival Anthropic is approaching $19 billion in annualized revenue.

An OpenAI IPO would be unlike anything the public markets have processed in years. The valuation, the opacity of the business model, the nonprofit-to-capped-profit structure, and the political entanglement — Greg Brockman and other executives have donated heavily to Trump-aligned super PACs — all make this a strange animal for traditional equity analysis.

One quarter. Three months. $297 billion into AI startups.

All of last year was a record at $425B. Q1 alone is on pace to nearly triple that.

At some point the question isn’t “is AI overfunded” — it’s “what happens when a meaningful fraction of these bets don’t return.” The answer is probably: a lot of people lose a lot of money, a few technologies stick around, and the narrative shifts to “the real AI” coming next.

Engineers built a memory device that keeps functioning at temperatures hotter than molten lava. 700°C. Normal chips fail around 125°C.

Where would you even use this? Inside jet engines. Deep geothermal sensors. Spacecraft re-entry. Places where electronics currently just die.

Not going in your phone but an interesting piece of engineering.

Silicon-carbon battery design. The energy density has always been there — the problem was stability, they’d degrade fast. New design addresses that.

If it performs outside the lab the way it does inside: Apple and Samsung could ship phones that last significantly longer. Not a little longer. Significantly.

The lab-to-product gap is real and often long. But this one has the right people paying attention.

Two research teams published findings saying we’re closer than anyone expected. One estimate puts the current largest quantum machine at more than halfway to the size needed to crack the encryption that secures basically everything online.

“Just around the corner” is relative. But this moved from “eventual theoretical problem” to “timeline question” faster than most people were tracking.

What happens to all the data that’s been harvested now, to be decrypted later? That’s the uncomfortable part of this.

Strip away the personal animosity and the lawsuit between Elon Musk and OpenAI is a fight about something that will define the AI industry for a decade: can a nonprofit that controls a powerful technology convert itself into a for-profit without betraying its founding mission?

Musk’s core legal argument is that he donated money and resources to OpenAI on the explicit basis that it was a nonprofit pursuing AI for humanity’s benefit. The conversion to a capped-profit structure — and the ongoing push toward a full for-profit entity — violates those terms, he argues. OpenAI counters that the mission has not changed, only the structure needed to raise the capital required to remain competitive.

The space industry before SpaceX and the space industry after SpaceX are different industries. Understanding what changed explains why the next decade looks nothing like the previous fifty years.

The core innovation was reusability. Rockets before SpaceX were expendable — you built them, launched them once, and they fell into the ocean. The cost of reaching orbit was priced accordingly: tens of thousands of dollars per kilogram of payload. SpaceX’s Falcon 9, which lands its first stage and reuses it across dozens of flights, collapsed that cost by roughly 90%. When launch becomes cheap, everything downstream changes.

Everyone knows Nvidia makes the chips that power AI. Fewer people understand why competitors have been unable to close the gap despite years of trying and billions of dollars of investment.

The hardware advantage is real but secondary. AMD makes competitive GPUs. Google has TPUs. Amazon and Microsoft have custom silicon. On raw performance for certain workloads, these alternatives are credible. The reason Nvidia keeps winning is not the chip — it is CUDA.

The TikTok situation in the US has become so legally and politically tangled that a clear summary is genuinely useful.

Congress passed a law requiring ByteDance, TikTok’s Chinese parent company, to divest its US operations or face a ban. The law survived a Supreme Court challenge. The deadline passed. TikTok went dark briefly in the US, then came back when the incoming Trump administration signaled it would not enforce the deadline immediately, seeking instead a negotiated outcome. The app has been operating in a legal grey zone since.

Self-driving cars have been “two years away” for almost fifteen years. Something has genuinely changed in the last eighteen months. Understanding what still needs to happen is more useful than the hype in either direction.

The technology itself has cleared important thresholds. Waymo’s fully autonomous robotaxi service is operating commercially in multiple US cities with safety records that compare favorably to human drivers. Tesla’s Full Self-Driving software handles an increasingly wide range of scenarios without intervention. The question has shifted from “can this be done” to “can this scale.”

Nuclear power was supposed to be a fading technology. Expensive, politically toxic after Fukushima, outcompeted by renewables. The reversal now underway is genuine and worth understanding.

The driver is AI. Data centers powering large language models and the infrastructure they require consume enormous and rapidly growing amounts of electricity. Unlike residential or commercial demand, these loads are constant — 24 hours a day, 365 days a year. Solar generates during the day. Wind generates when the wind blows. Nuclear generates all the time, regardless of conditions. For a tech industry trying to guarantee power availability at scale, nuclear has become newly attractive precisely because of the attribute that made it economically awkward in a grid context: it does not stop.

The debate about social media and mental health has been running for a decade. The research has caught up, and the picture is sharper than it used to be.

The harm is not social media use broadly. It is specific: heavy algorithmic feed consumption, particularly among adolescent girls, correlates meaningfully with depression, anxiety, and disordered eating. The correlation survives controls for pre-existing conditions and reverse causality in the most rigorous studies now available. It is not a proven causal chain in every case, but it is strong enough that the “no evidence of harm” position is no longer defensible.