The day's tech, sifted: Aug 15, 2026

What matters today: Anthropic told prospective investors it generated $11.5B or more in Q2 revenue with positive adjusted operating income, up from $787M a year earlier and $4.73B in Q1, and projects 2028 revenue of roughly $190B to $200B against a $47B run rate in May as bankers price the company ahead of an IPO, the same day it raised its misalignment risk estimate from very low to low and said it won't release a stronger internal model called Model 2. OpenAI's own IPO run looked messier: Financial Times sources say repeated executive reshuffles and departures have frustrated staff, and the company disbanded its "preparedness" safety team in July. The Chinese open-weight chase kept pace too: Z.ai says GLM-5.3 scored 84.5 percent on the CyberGym benchmark against Mythos 5's 83.8 percent, a 750B-parameter model Interconnects notes reached the frontier through post-training scale rather than distillation. Elsewhere, a macOS screen-sharing flaw is under active cryptomining exploitation, PayPal is fielding a buyout approach from Stripe and Advent above $60.50 a share, and Nvidia disclosed nearly $51B in combined SpaceX and Intel stakes while reworking its OpenAI data-center financing.

AI / LLMs

• Anthropic told investors it generated $11.5B+ in Q2 revenue with positive adjusted operating income, a 14-fold jump from $787M in Q2 2025 and up from $4.73B in Q1 2026, while projecting 2028 revenue near $190B to $200B against a $47B run rate as of May: the same week it raised its own misalignment risk estimate from very low to low and confirmed it won't release a more capable internal model dubbed Model 2, pairing hypergrowth with a rare admission of rising risk.
• Anthropic detailed how the text watermark it said this week it would roll out to meet the EU AI Act's marking rule actually works: sparse in code and factual text, only signaling Claude was likely involved rather than proving it, and fading after a full rewrite.
• OpenAI's repeated executive reshuffles and departures have frustrated staff preparing for an IPO, and the company disbanded its "preparedness" safety team in July, Financial Times sources say: the churn extends a pattern that surfaced with Brad Lightcap's exit last week and widened as Greg Brockman moved into what insiders call "founder mode" to rebuild the leadership bench.
• Z.ai says its open GLM-5.3 scored 84.5 percent on the CyberGym cybersecurity benchmark against Mythos 5's 83.8 percent, keeping its most sensitive functions gated to verified users; Interconnects credits the 750B-parameter model, a third of Kimi K3's size, to heavier post-training rather than distillation from US labs, part of why Chinese frontier releases keep landing within days instead of the months US labs spend on safety testing before shipping.
• Alibaba's open-weight models have logged 3B+ global downloads over the past six months, more than Google's 418M and Meta's 227M combined for 2026 so far, per Hugging Face data cited by Bloomberg: adoption, not just benchmark scores, has tilted toward Chinese open models.
• DeepSeek quietly shipped V4 Pro 0813 with a 5x leap in agentic coding and open weights under an MIT license, paired with a 264 percent price increase landing August 16: a rare price hike inside a week otherwise dominated by cuts.
• METR raised $71M in six months to scale independent, third-party stress-testing of frontier AI models, funding an outside check on exactly the kind of self-reported risk numbers Anthropic published today.

Devtools & Infra

• Dynatrace agreed to acquire Arize, an AI observability and AI-development-lifecycle startup, for $915M including about $815M in cash: another AI-native tooling company getting absorbed into an incumbent observability platform rather than building the category out on its own.

Security & Privacy

• A macOS screen-sharing vulnerability tracked as CVE-2026-65400 is under active exploitation, Dutch officials warn: attackers with root access on systems exposing port 5900 have been dropping Monero cryptominers, and while Apple patched it last week for Tahoe, Sequoia and Sonoma, machines with screen sharing still exposed to the internet remain at risk.
• A Connecticut judge caught what appears to be the first US case of a plaintiff hiding AI-only prompt injections inside court filings: invisible text instructed any AI reviewing the document to side with his arguments and ignore prior denials. The judge ruled it had no effect on the case's merits but called the tactic dangerous and one courts will likely see again.

Startups & Industry

• PayPal is in talks to sell itself to a group including Stripe and private-equity firm Advent, which offered $60.50 a share in July: PayPal is holding out for a higher price, a striking role reversal for a payments pioneer now being circled by the fintech company it once dwarfed.
• Nvidia disclosed a nearly $21B stake in SpaceX and a $30B stake in Intel, on top of up to $2B in xAI and $5B more in Intel from 2025, the same day it reworked its OpenAI Ohio data-center financing to initially guarantee only half of its planned $250B backstop: Nvidia's balance sheet is now wired into nearly every major AI bet at once, the kind of financial engineering Stratechery's Ben Thompson flagged this week as expanding the AI bubble's blast radius even as it buys the industry time.
• Databricks raised another $5B, its second megaround in eight months, valuing the company at $190B and up from $134B in December, on a revenue run rate that's now topped $7B; the same week Igor Babuschkin's months-old River AI raised $1.1B in seed and Series A funding with backing from both Nvidia and AMD Ventures.
• Waymo won approval from the California Public Utilities Commission to expand its robotaxi service across 18 counties, including Sacramento and San Diego, its biggest expansion yet.
• Commerce Secretary Howard Lutnick said the US is "not in favor" of Apple buying Chinese memory chips, telling the company to find "other solutions to the memory issue": a public nudge against Apple's most obvious escape from a memory crunch Washington would rather it solve at home.
• The US plans to tell 35 partner countries they'll be excluded from its Pax Silica AI initiative if they also join China's rival framework, according to a source and letter: an explicit pick-a-side ultimatum extending the chip and model rivalry into diplomacy.
• The US OCC conditionally approved a bank charter for the Trump family's World Liberty Financial, letting it directly issue its USD1 stablecoin: a national bank charter tied directly to a sitting president's family business, the clearest sign yet of the administration's stablecoin ambitions and its own finances intersecting.

Elsewhere

• A PBS affiliate in St. Louis is suing Iron Mountain Data Centers after its cloud storage provider went dark, seeking back 50TB of archival footage including its COVID-19 and 1993 flood coverage: most of the 70 years of material is, by the station's own account, unique and irreplaceable, a small-scale preview of what happens when a storage middleman simply disappears.
• The Verge's back-to-school laptop guide leads with a warning: RAM and storage prices are "out of whack" and likely to keep climbing through 2027, since next year's memory supply is already spoken for: the same memory crunch Lutnick just told Apple to solve without Chinese chips is now showing up in what a family pays for a student laptop.

Hacker News

The front page skewed AI: Qwen 3.8 27B landed on Hugging Face and pulled the day's biggest crowd, nearly 1,000 points and a 620 comment discussion arguing over the odd version jump and real world benchmarks. Google detailed homomorphic encryption for private AI inference, letting models compute on encrypted data without seeing plaintext, with commenters pressing on the latency cost. Mixedbread shipped Toast 1, a new embedding model, and one builder wrote up using Codex to auto research a 232x kernel speedup, read by many as a case study in agentic coding rather than a rigorous benchmark.

Elsewhere, Firefox is now the last major browser still shipping uBlock Origin support, reviving ad blocking as a real browser differentiator. A cryptographer's essay on "going dark" and law enforcement hacking drew heavy debate over lawful intercept versus device security. France's top court blocked a ban on social media for under 15s on free expression grounds. On the dev side, a RISC-V critique argued the ISA is repeating old mistakes, and a rant against oversized pull requests resonated with reviewers tired of unreviewable diffs.

Threads

• Anthropic's revenue trajectory (from $787M to $4.73B to $11.5B+ a quarter, with a $190B to $200B 2028 target) and Nvidia's reworked, ever-larger web of stakes and financing deals are the same story from two sides: the AI buildout's capital math is getting more elaborate, not simpler, exactly as Stratechery warned this week.
• OpenAI's disbanded safety team and widening executive churn landed the same day Anthropic raised its own misalignment risk estimate: one lab documenting more risk ahead of a listing, the other losing the staff meant to catch it.
• Chinese open models kept closing the gap from two directions at once: GLM-5.3 nearly matched Mythos 5 on a cybersecurity benchmark, DeepSeek V4 Pro claimed a 5x agentic coding leap, and Alibaba's models alone now out-download Google and Meta's combined, all landing the same week Alibaba and Z.ai's earlier releases dominated the price war.
• Anthropic raised its own misalignment risk estimate the same day METR raised fresh money specifically to stress-test frontier models independently: self-grading and outside verification both got more serious at once.
• The memory crunch squeezed from both ends: Lutnick publicly blocked Apple's most obvious fix (Chinese chips) just as RAM and storage costs pushed even a basic back-to-school laptop higher.
• Regulators handed out very different green lights the same day: Waymo got the state's blessing to expand across 18 counties, while the OCC handed a federal bank charter to a stablecoin business owned by a sitting president's family.