The companies racing to build frontier AI just agreed to referee themselves. Over the same week, launches, investor days, and earnings put a price on that arrangement. Microsoft is charging for finished work, NVIDIA is selling agent containment as infrastructure, and the suppliers AI can't route around, from Synopsys to Micron, are collecting the premium. Patrick Moorhead and Daniel Newman separate signal from stagecraft, then split hard on whether the White House accord actually binds these key players to anything. The handpicked topics for this week are: White House Gathers AI Leaders Behind Voluntary Safeguards: Executives from NVIDIA, AMD, OpenAI, Anthropic, SpaceX, Microsoft, Google, and Amazon met at the White House to agree on a safety framework after a string of agent containment failures. Daniel Newman described two camps. The labs asking to be regulated are, in his view, partly chasing regulatory capture and positioning for multi-trillion-dollar IPOs, while Jensen Huang's camp argues the risk can be managed technologically. The group landed on immediate self-regulation, peer evaluation of models, and greater board oversight, with Mark Zuckerberg reading out the outcome. Newman called it a productive meeting with the right people in the room, since pausing or stopping AI isn't a realistic option. (The Decode) NVIDIA Puts Agents in a Managed Sandbox: In the same week, NVIDIA introduced its open agent safety platform, which combines a sandboxed agent runtime with a reference design and can use BlueField-4 for enforcement outside the host. Patrick Moorhead highlighted Hugging Face CEO Clem Delangue's claim that OpenAI could not have hacked Hugging Face if it had been running the platform. Moorhead credited NVIDIA's leadership while noting the framework is open, runs on all x86, including Intel and AMD, and can integrate DPUs from AMD and the hyperscalers. He expects AWS, which has not yet signed on, to arrive with an approach that spans GPUs and Trainium. (The Decode) Microsoft Resets Copilot Around Work Completed: CEO Satya Nadella positioned the new Copilot as the operating system of business, with a Home tab that combines chat and cowork, a Code tab that builds apps and workflows from plain language, and Autopilot for persistent background agents. Patrick Moorhead, who was briefed by the product team, believes this version has a high chance of sticking, flagging a pricing model that charges for work done instead of per seat. He named governance, productivity data, and Microsoft Fabric as Microsoft's moat. Daniel Newman cited an ETR pulse of 50 Fortune company CIOs in which about 70% want AI's pace slowed because they can't keep up. He argued Microsoft should be the adult in the room that sets the enterprise pace, the same play ServiceNow is running with Flow by wrapping vibe-coded tools in governance and SLAs. (The Decode) Synopsys Brings OpenAI and Amazon Into AI-Native Chip Design: Synopsys launched GPT Synopsys, a frontier model co-designed with OpenAI and layered on Synopsys' full tool chain. Daniel Newman read the deal as OpenAI conceding that EDA tooling, IP, and experience are essential to cutting chip design cycles to six months or less. Amazon became a lead customer for Synopsys' application-optimized silicon IP and expanded its EDA and agentic engineering relationship, and Synopsys launched long-horizon Agent Engineer agents with Autopilot orchestration, citing 40 to 50 engagements already underway. Patrick Moorhead shared Synopsys chief product officer Ravi Subramanian's framing that AI reasons and explores while Synopsys generates and validates, along with the customer line "In AI, we believe, but in physics, we trust." Moorhead also warned that zero data retention doesn't mean companies aren't sharing IP, and pointed to confidential computing as the next shoe to drop. (The Decode) Anthropic, OpenAI, and Google Match Prices on New Models: Claude Sonnet 5.5, GPT-6.1 Sol, and Gemini 4 Argon all list at $2 input and $10 output, with Argon still in security testing. Patrick Moorhead argued the price that matters is the cost of a known good output. He cited Signal65 findings that OpenAI's $200 Pro plan now delivers less output for the same price, and that GPT-6.1 Sol at max effort raised quality without lowering the bill per correct task. He compared the labs' efficiency push to Chinese open model developers, citing DeepSeek's 85% KV cache reduction. Daniel Newman called Signal65's Pinnacle testing the most important resource for enterprises trying to understand the token economics of running agents in production. (The Decode) AMD Makes an $8.2 Billion World Labs Bet: AMD agreed to acquire Fei-Fei Li's World Labs for $8.2 billion, about 1.6 times the $5 billion valuation Bloomberg reported. Daniel Newman called it a talent grab and a bet on physical AI and world models, made possible by AMD's trillion-dollar market value, though the deal came with almost no disclosed financials. After a Q&A with AMD AI chief Vamsi Boppana, Patrick Moorhead described it as a talent and model buy with little near-term revenue, scoped across content, media, robotics, embedded, and data center. He said the team will need to keep building models to retain talent, and could give AMD sharper insight into its hardware roadmap. (The Decode) The Flip: Will the White House AI Accord Solve Real Problems?: Patrick Moorhead argued yes. Getting competitors behind common responsibilities makes safety hard to dismiss, independent evaluation and board-level review offer concrete mechanisms, behavior is already shifting with OpenAI pausing frontier training after its containment failures, and the FTC's existing investigation provides a backstop. Daniel Newman countered that the accord's "believe" and "should" language reads like a mission statement, the pledge is voluntary with no enforcement, the floated oversight committee could come from the signatories themselves, and OpenAI and Anthropic have already committed to third-party evaluators. He pointed to reports of Meta's Muse agent exposing users' private data the same week, and said companies playing offense on safety ship evidence while companies playing defense ship photo ops. Note: The Flip assigns Patrick Moorhead and Daniel Newman opposing sides of a simulated debate, positions argued are not necessarily their own views. (The Flip) Synopsys ($SNPS) Raises Guidance and Adds a $1 Billion Buyback: Synopsys shares gained 10% to 12% after Investor Day as the company raised revenue guidance from about $10.81 billion to nearly $11.2 billion, lifted margin guidance, and added a $1 billion buyback. Daniel Newman points to the systems side of the business, spanning physical AI, robotics, and simulation. Patrick Moorhead suggests Synopsys' terminal value is underscoped in a market of "two and a half players," referring to Synopsys, Cadence, and to an extent, Siemens. He expects AI to bring more customers into chip design, gated mainly by leading-node capacity at TSMC, with Samsung and Intel adding capacity. (Bulls and Bears) HPE ($HPE) Lifts Its Networking Outlook at Investor Day: Four weeks after its September 2 earnings call, HPE raised its 2027 networking guide from a 14-to-17 range to the high teens and low 20s, which Patrick Moorhead tied to AI monetization and Helios inside AMD's rack. Networking margins hit a record 40% in Q3, and a $1.2 billion Helios rack order arrived, business Moorhead called high margin since HPE passes on lower-margin neocloud deals. Daniel Newman noted integration savings nearing $1 billion in annual run rate and a stock that has roughly tripled in a year. Moorhead argued HPE should be modeled more like Cisco than Dell. (Bulls and Bears) Micron ($MU) Posts 87% Margins: Micron delivered $54 billion in revenue and $32.87 in EPS at about 87% margins, and guided to $61 billion, $38 a share, and 86% to 87% margins. Daniel Newman noted the growth comes almost entirely from pricing, with supply constrained in 2027 and likely 2028, and said roughly $70 billion in cash could turn Micron into a major ecosystem investor. Patrick Moorhead added that customers put down $12 billion in deposits to fund clean rooms arriving in late 2028, and that Goldman, Rosenblatt, D.A. Davidson, and Mizuho all raised targets, led by Rosenblatt's move from $1,500 to $1,900. Newman added that some accelerator makers are returning wafer allocations because they can't secure memory, capacity he expects NVIDIA to absorb. (Bulls and Bears) Accenture ($ACN) Jumps 16%: Accenture rallied 16% in a day, though Patrick Moorhead noted half of EPS growth came from a lower share count after $2.3 billion in buybacks for the quarter and $7.5 billion for the year. Fixed-price work, including outcome-based contracts, now makes up 65% of bookings, which ties margin to delivery instead of the rate card. Moorhead doesn't think Accenture is out of the woods yet with roughly a million employees. Daniel Newman argued the "AI kills services" call has been wrong so far, because enterprises want to deploy agents but lack the resources to do it safely and securely, and they turn to partners like Accenture to get it done. (Bulls and Bears) Watch the full episode at sixfivemedia.com and subscribe to our YouTube channel so you never miss an episode. The Decode White House AI Safeguards https://www.reuters.com/business/ftc-opens-probe-into-ai-giants-including-anthropic-openai-new-york-post-reports-2026-09-30/ Microsoft's New Copilot https://blogs.microsoft.com/blog/2026/09/25/introducing-the-new-copilot-with-home-code-and-autopilot/ Synopsys, OpenAI, and Amazon https://x.com/danielnewmanUV/status/2105363539837829130 Claude Sonnet 5.5, GPT-6.1 Sol, and Gemini 4 Argon https://www.anthropic.com/claude-sonnet-5-5 https://openai.com/index/introducing-gpt-6-1-sol/ https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ https://pinnac
