Anthropic's IPO Gamble, Apple's Agent Lockdown, and a Chip Bust

Anthropic Presses Ahead With Thanksgiving IPO Despite Losses and Scrutiny

Anthropic is still aiming for a pre-Thanksgiving public offering even after a leaked filing showed an $8 billion operating loss. Big Technology reports that revenue could jump from $400 million in 2024 to $4.6 billion in 2025, but says the company spent $7.3 billion on infrastructure last year alone and expects to spend $518 billion over the next decade, with about 80 percent of that non-cancellable. Sources: SiliconANGLE - AI, Big Technology

The stakes reach beyond one company. Goldman Sachs says the S&P 500 is showing its worst breadth since the dot-com bubble, with the median stock 16 percent below its 52-week high even as the index is up roughly 13 percent this year, and Goldman estimates half of the index's earnings-per-share growth is coming from AI investment. Nvidia alone added more than $1 trillion in market cap in 2026. Source: Big Technology

The landscape working against the IPO keeps growing. Meta's cheaper Muse model is good enough that SAP's CEO told Big Technology, "We don't need to always use the best, best model from an outcome perspective," the FTC is reportedly investigating both OpenAI and Anthropic over consumer risk, and OpenAI dismissed three safety researchers accused of sharing confidential material. Oura postponed its own IPO, Accelevation's debut struggled, OpenAI has signaled it might wait until next year, and the Financial Times said the build-up to Anthropic’s stock market debut has felt "discordant, even creepy." Florida governor Ron DeSantis added to the political noise, posting that "Superintelligence factories = surveillance centers" right after AI CEOs met with Trump. Sources: SiliconANGLE - AI, Big Technology, Financial Times - AI

Apple Tightens Mac Privacy Controls After Meta Agent Scare

Apple is changing how macOS handles Full Disk Access after columnist Jason Aten said Meta's agent Muse referenced a private Apple Messages thread he never authorized it to read. Meta CTO David Singleton denied this was possible without a user opting into both full disk access and a separate Messages connector, but security researcher Patrick Wardle countered that with full disk access, virtually any non-root file, including browsing history and chats, is readable. Sources: Ars Technica - AI, TechCrunch - AI

Apple did not name Meta or Muse, but it warned that "as AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially," and said granting that level of access will now require "very explicit user action." Nobody has proven Muse actually read the messages without permission, but Apple said the setting itself created risks that users might not fully understand. Sources: Ars Technica - AI, TechCrunch - AI

California Man Charged in $300 Million Nvidia Chip Smuggling Case

Federal prosecutors have charged a California man with smuggling more than $300 million worth of computer servers containing export-controlled Nvidia GPUs into China, allegedly using false paperwork and shipments routed through third countries. He faces up to 20 years in prison if convicted on the most serious single counts. Sources: The Guardian - AI, Ars Technica - AI

The case is not isolated. Ars Technica notes an Nvidia manager in Taiwan was previously linked to a similar smuggling scheme tied to former Supermicro staff, highlighting ongoing attempts to route controlled chips to China through intermediary countries. Sources: The Guardian - AI, Ars Technica - AI

Nvidia Adds a Smaller, Cheaper DGX Spark for Local AI Agents

Nvidia is rolling out a 64 gigabyte version of its DGX Spark desktop AI system this month through manufacturer partners including Acer, ASUS, Dell, Gigabyte, HP and MSI, alongside the existing 128 gigabyte model. It keeps the same GB10 Grace Blackwell Superchip and NVIDIA AI software stack as the 128GB version, and Nvidia says it can run capable local agents on device without cloud dependency and support models and agentic applications at practical scales fully on device. Sources: NVIDIA blog, MarkTechPost

Suno Adds Narrated Speech, but Copyright Cloud Lingers

Suno has launched Speech, a public beta feature that turns a script or voice description into spoken audio paired with matching background music in a single track up to about eight minutes long. Product chief Jack Brody says the company tested it with a small group for a month and is candid about its rough edges, joking that "occasionally, British accents can wander off to Australia and back." Sources: The Decoder, The Verge - AI

The launch comes as Suno still has not disclosed what data trained the model, major record labels are suing the company, and a Munich court recently rejected fair use as a defense for using copyrighted training data. The tool works today, but a real legal cloud hangs over it. Source: The Decoder

Amazon Pledges $1 Billion for Data Center Communities Amid Backlash

Environmental nonprofit Stand.Earth called the plan "corporate propaganda," noting it does not mention the power plant Amazon is backing in Pecos, Texas, which the group expects to become the single largest source of climate pollution in the country, producing 33 million tons of air pollution a year. Source: Ars Technica - AI

Open-Source Model Writes Cited Science Reports Faster, With a Caveat

Researchers behind the Asta platform have open-sourced AstaBrief, an eight billion parameter model built from Qwen3-8B that turns a research question and literature excerpts into a cited report. Across their pipeline it averages 51.1 seconds per report in Fast mode versus 178.5 seconds for a Claude-powered Thinking mode, about three and a half times faster, and the team is releasing both weights and training data. Source: Hugging Face blog

The catch is that most of the quality evaluation was done in 2025 against the proprietary models available then, and the researchers acknowledge they have not rerun the comparison against today's frontier models. The speed advantage is real and documented, but the quality comparison is already a year stale by their own admission. Source: Hugging Face blog

Researchers Estimate Compute Could Soon Run Hundreds of Millions of AI Agents

Epoch AI estimates that compute coming online over the next two years could support tens to hundreds of millions of AI agents running today's most capable models continuously, or billions running cheaper ones. Using projected high-bandwidth memory shipments through 2027, the group calculates those top-tier agents running nonstop could supply as many weekly working hours as roughly 140 million to 700 million full-time employees, compared with about 100 million knowledge workers in the entire United States. Source: Epoch AI

With a more efficient model like DeepSeek V4 Pro, the same hardware could run about 1.9 billion concurrent agents, matching the working hours of 8 billion full-time workers. Epoch's revenue math suggests that if AI company revenue keeps growing fivefold a year it reaches roughly $1 trillion by the end of 2027, but if just 20 percent of that compute serves paid inference at today's prices, the implied revenue is $2.6 trillion to $5.3 trillion annually. The estimate assumes demand and pricing hold steady even though API prices have been falling 50 percent a quarter. Source: Epoch AI