Afternoon Brief · September 2nd 2026

Afternoon Brief, September 2nd 2026

Broadcom beat estimates and still sank the stock on its . Google shipped Gemini 3.8 Flash, plus a cyber-only sibling gated behind a new vetting program. Wall Street snapped a two-day slide as Nvidia jumped on the report. And the Justice Department told a federal court that training on the news is .

10 min read·Editorial by Nour Haddad

NHNour Haddad
Deployment & the Floor

Updated

Broadcom's AI semiconductor revenue hit a record, and the stock fell anyway

Broadcom reported fiscal third-quarter results after today's close: $29.59 billion of revenue against the $29.47 billion Wall Street wanted, adjusted earnings of $3.32 a share against a $3.22 estimate, and net income more than tripling to $13.09 billion. AI semiconductor revenue reached $16.7 billion, up more than 220% from a year earlier, another record for the segment. What moved the stock was the number that did not rise: fiscal fourth-quarter revenue guidance of $34.8 billion, short of the roughly $35 billion analysts wanted, with no lift to the full-year AI semiconductor forecast. Shares closed the regular session at $367.24, up 23% over the past year, then fell as much as 3.6% after hours. A beat that still sinks the stock says the AI-capex trade is now pricing acceleration every quarter, and a record quarter that merely repeats its own growth rate reads, to that market, as deceleration. SourcesA B

Broadcom, quarter ended August 2026

Revenue and EPS beat estimates. Q4 guidance of $34.8B came in below the ~$35B analysts wanted, and shares fell as much as 3.6% after hours.

Updated

Google ships Gemini 3.8 Flash, and a cyber-only sibling gated behind a vetting program

Google DeepMind released Gemini 3.8 Flash on Wednesday, the third Flash model in six weeks, generally available through Google AI Studio and the Gemini Enterprise Agent Platform. It scores 90.8% on Terminal-Bench 2.1, up from 81.6% for 3.7 Flash, and Google says it holds Flash-tier pricing while beating larger on long-horizon coding tasks. Alongside it, Gemini 3.8 Flash Cyber: a variant tuned for autonomous vulnerability discovery and automated patching, restricted to a new Fairwind Program that vets governments, critical-infrastructure operators and software maintainers before granting access. Google says the cyber variant leads the CyberGym Pass@1 at 86.2%, ahead of Anthropic's Mythos 5 and OpenAI's GPT-5.5-Cyber. Three labs have now converged on the same structure within a week: ship the capable general model to everyone, and put the version tuned for offense behind a credential. SourcesA B

Wall Street snaps its losing streak as Nvidia jumps on the Hugging Face report

The S&P 500 rose 0.46% to 7,666.60, the Nasdaq gained 0.45% to 26,217.83 and the Dow added 0.56% to 53,061.95 on Wednesday, ending a two-session slide tied to this week's Iran strikes. Nvidia led the rebound, up 3.2% to $224.41, part relief and part Bloomberg's report that a Hugging Face deal could close as soon as this week. Bond markets told a different story: the 10-year Treasury yield climbed to 4.814%, its highest since November 2023, and crude held near $90.63 a barrel. Stocks rising alongside a multi-year high in yields is a bet the AI trade can keep clearing a higher cost of capital, a bet Broadcom's after-hours guidance complicated within hours of the close. SourcesB B

The Justice Department tells a federal court that training on the news is fair use

The Justice Department filed a Wednesday in the New York Times' copyright suit against OpenAI and Microsoft, backing OpenAI's fair-use defense. The filing argues a "robust" domestic AI industry serves national security, economic development and scientific advancement, and that those interests should weigh in favor of treating AI training on copyrighted material, including news articles, as fair use. It does not bind Judge Sidney Stein, who is hearing the case in the Southern District of New York, though it puts the executive branch on record backing OpenAI's side of the industry's highest-profile copyright fight, three years after the Times sued. The Times called it "siding with a handful of trillion-dollar AI companies at the expense of the countless American creators whose work they stole." The brief is a policy argument about what the country needs, layered on top of a legal one about transformation and market harm that Judge Stein still has to decide on its own terms. SourcesB B

Anthropic lets big customers keep misuse-monitoring data out of Anthropic's hands

Anthropic launched Enterprise Frontier Safeguards on Wednesday, combining with automated misuse detection that runs against storage the customer owns instead of storage Anthropic owns: activity logs can live in a customer's Amazon S3, Azure Blob or Google Cloud Storage account, under keys and audit logs the customer controls, and flags from the automated system route straight to the customer's own review team. The monitoring watches for the categories a frontier lab worries about most, attempts at offensive cyber or biological capability and signs of stolen credentials, across a rolling window of traffic. Support is rolling out across Claude Code, Claude Enterprise, the Claude Platform, Bedrock, Google's Agent Platform and Microsoft Foundry, at no charge beyond the customer's own cloud storage bill. It is a second kind of gate from this week's model releases, a wall around who gets to watch the traffic rather than around the capability itself, and watching your own traffic is a benefit sized for institutions large enough to staff a review team in the first place. SourcesA B

CrowdStrike lets one AI attack a company so another AI can defend it

CrowdStrike introduced SafeMind on Tuesday at its Fal.Con conference in Las Vegas, a pair of models built with Nvidia: Red Tempest, an offensive model trained partly on 15 years of CrowdStrike's own incident-response data, and Blue Solano, a defensive model that remediates what Red Tempest finds. Red Tempest probes a of a customer's environment, built inside Nvidia's simulation stack, for attack paths; Blue Solano closes them; the cycle repeats until none are left. Nvidia's Jensen Huang and CrowdStrike's George Kurtz announced it together on stage in front of roughly 10,000 security professionals. SafeMind runs inside the Falcon platform, with standalone access through a program CrowdStrike calls Project QuiltWorks. It is the AIR Security pitch from this morning's brief run in reverse: instead of filtering what reaches an , CrowdStrike is manufacturing the attacker itself, on a leash, to find the hole before someone without one does. SourcesA B

ChatGPT gets a read-only pipe into Epic's patient records

OpenAI said this week that healthcare organizations can connect their Epic electronic health record environments to ChatGPT for Healthcare, letting clinicians pull a patient's notes, labs, medications and specialist documents into ChatGPT to prepare for an appointment, or reach ChatGPT directly inside the EHR workflow without leaving the chart. Access is read-only; nothing writes back to the record. OpenAI says physicians rated 99.1% of ChatGPT's responses safe across 4,363 ratings spanning 27 clinical use cases tested with connected EHR context, and UCSF Health is the pilot partner. A companion Healthcare Public Data plugin adds structured access to PubMed, DailyMed and CMS coverage data. Epic's system serves more than 325 million patient records, which is the actual news here: the integration is less a new capability than a distribution deal, putting a general-purpose chatbot in front of the record system most of American medicine already runs on. SourcesA B

Sources

33 citations · 11 primary · 22 secondary

Morning Brief, September 2nd 2026