Windows Goes Hybrid, Marvell’s $90B Target, and the AI Routing Race | The Six Five Pod Ep. 323

On Episode 323 of The Six Five Pod, Patrick Moorhead and Daniel Newman unpack the biggest stories from the week, including Microsoft and NVIDIA’s hybrid intelligence push for Windows, Marvell’s Investor Day, new models from Anthropic, Mistral, and Reflection, the AI labs’ accountability hearings in New York and Australia, Google’s model-neutral Gemini agent, and whether temporary construction jobs justify AI data centers. Plus, Bulls & Bears covers Thursday’s chip-led pullback, NVIDIA’s run past $240, and Samsung’s record Q3 guidance.

The handpicked topics for this week are:

1. Microsoft and NVIDIA Bring Governed AI Agents to the PC: Satya Nadella, Jensen Huang, and Pavan Davuluri shared the stage on October 7 to position Windows as a hybrid intelligence platform that routes AI work between local models and the cloud. The launch included the Surface Laptop Ultra powered by NVIDIA RTX Spark, general availability of Microsoft Execution Containers to restrict what agents can access, and DGX Station for Windows, with systems from ASUS, Dell, HP, Lenovo, and MSI shipping October 16. Pat calls it the highlight of his week because it brings together local models, governed subagents, and the lowest-cost placement of every workload. Pat argues local AI pays off if developers can shift a meaningful share of their tokens onto the device, and it also keeps proprietary data away from frontier labs. He wants to build an end-to-end routing test in Signal65 Pinnacle. Dan frames it as pure math: match each task to the intelligence it actually needs, and the cheapest compute that can run it, and wrap it in tooling simple enough to become a managed service. (The Decode)

2. Marvell Puts Connectivity, Not Just Custom Chips, at the Center of AI: At its October 6 Investor Day, Marvell set a bottoms-up fiscal 2031 revenue target of $70 billion to $90 billion and raised its fiscal 2028 outlook to about $20 billion. Dan calls it one of the best-orchestrated investor days he has seen, with Matt Murphy finally making the call on growth after years of defending the custom silicon business. Dan argues Marvell made the right move leading with XPU attach connectivity across scale-up, scale-out, and scale-across, with custom work at all four major hyperscalers. Pat highlights the new “scale-in” category, connectivity inside the package, and notes the target is roughly a 10x from where Marvell stood a few years ago. Pat is puzzled that the stock faded from about a 10% gain to roughly 4%. Dan points to execution risk, a market hypersensitive to AI demand headlines, and a buy side that had already modeled much of the growth. (The Decode)

3. Anthropic, Mistral, and Reflection Compete on Cost and Useful Work: Anthropic released Claude Haiku 5.5 on October 7, which Anthropic says runs about 75% cheaper per task on average than Haiku 4.5 and targets summaries, classification, database queries, and coding subagents. Mistral launched Mistral Large 4, which Pat says drew more excitement than any Mistral model before it. Reflection unveiled Beam on October 5, a 501-billion-parameter open-weight model with 23 billion active parameters, positioned against Z.ai’s GLM-5.2 with far less inference compute required. Pat wants both Mistral and Reflection running in Signal65 Pinnacle. He also calls out Anthropic for merging chat and work modes in Claude, which he says pushed one of his show-notes runs from 22 minutes to nearly three hours with no better output. Dan warns that users who aren't paying attention will burn through token limits faster on lower-value work. (The Decode)

4. OpenAI and Rival Labs Face Accountability Tests in New York and Australia: OpenAI, Anthropic, Google, and Meta testified under oath before the New York City Council on October 5, pressed on liability, independent testing, and shutdown capability. In Sydney, OpenAI Chief Strategy Officer Jason Kwon apologized to Australia's parliamentary inquiry after an OpenAI agent breached a Medicare statistics portal, and acknowledged notification should have come sooner. Dan sees more posturing than progress and warns that bipartisan anti-AI legislation could slow the data center buildout he views as critical to US GDP growth. He points to a new federal draft proposing developer liability for harmful agent conduct. Pat frames it as accelerators versus decelerators, and he argues that agent escapes trace back to human engineering choices like skipping sandboxing, not to AI itself. (The Decode)

5. Google Previews a Workplace Agent That Can Use Claude, Not Just Gemini: At Gemini at Work 2026 on October 8, Google Cloud introduced the Gemini agent, which keeps persistent context across Workspace, Microsoft 365, and Slack and keeps working through subagents after the user closes the laptop. The agent routes jobs across Gemini and Anthropic's Claude models today, with private and open models planned, alongside Agent Sandbox and Agent Gateway for enforcement. Pat calls model choice exactly what enterprises want and stresses how hard smart routing is to do at microsecond speed. He is skeptical until he can test it, noting his personal Gemini account still outpaces his paid Workspace account on features, memory, and models, and that Gemini Spark has yet to reach Workspace. Dan sees Google and Microsoft as the companies best positioned to be the adults in the room for the enterprise AI canvas, but says paying customers should get the best access, not less. (The Decode)

6. Are temporary construction jobs a strong reason to approve AI data centers? Daniel takes the FOR side: a single data center can require 1,000 to 10,000 construction workers, apprentices earn $40,000 to $60,000 a year, and experienced electricians clear six figures without a four-year degree. With roughly 20,000 electricians retiring each year and 30% of union electricians between 50 and 70, he calls the buildout the largest apprenticeship engine in decades, and argues communities should bargain hard rather than walk away. Patrick takes the AGAINST side, citing a Virginia audit that found roughly 50 ongoing jobs for every 1,500 workers at peak construction. Pat adds that statewide estimates showed rising electricity costs outlasting the construction wages, and argues every jobs promise should become a measurable obligation before approval. After the debate’s conclusion, Pat admitted he agrees with Dan. (The Flip)

7. Chips Lead Thursday's Pullback While Software Holds Up Better: A Financial Times report questioning how OpenAI's revenue is counted, gross versus net, triggered a sharp chip selloff, with money rotating into software and energy. Pat says the FT should be embarrassed for conflating ARR and revenue, and he was surprised chip stocks had not recovered by Friday. Dan attributes much of the move to algorithmic selling that reacts to trusted headlines and cascades as large blocks sell. Rising bond yields, open models pressuring frontier pricing, and questions about the fungibility of lab compute commitments add to the volatility. As Dan puts it, markets take the escalator up and the elevator down. (Bulls & Bears)

8. NVIDIA Crosses $240, Then Gives Back the Milestone: NVIDIA hit an all-time high above $240 this week, approaching a $6 trillion market cap, before Thursday's selloff. Dan argues NVIDIA should be a $10 trillion company within a couple of years if it keeps executing, and calls it cheap on price-to-earnings and growth. He contrasts how the market prices NVIDIA for any possible mistake while rewarding AMD, up roughly sixfold in a year, as if it carries no execution risk. Pat is waiting for NVIDIA’s next breakout as revenue heads toward $500 billion and beyond, and congratulates the bag holders. (Bulls & Bears)

9. Samsung's Preliminary Earnings Set the Next Semiconductor Checkpoint: Samsung guided to roughly 107.4 trillion won in Q3 operating profit, about $80 billion, on approximately 195 trillion won in sales, the first time a tech company has topped 100 trillion won in a single quarter. Pat notes Samsung missed on revenue but still posted NVIDIA-sized operating profit. He shares rumors that Samsung's HBM4E is the only HBM passing Vera Rubin power tests, possibly because Samsung builds its own base die while SK hynix and Micron outsource theirs to TSMC. Pat also flags Lisa Su's openness to other foundries as a potential opening for Samsung. Dan says the memory boom has transformed Samsung from written off to one of the most capital-rich players in semiconductors. (Bulls & Bears)

Watch the full video at sixfivemedia.com, and subscribe to our YouTube channel so you never miss an episode.

The Decode

Microsoft and NVIDIA Bring Hybrid Intelligence to Windows; Surface Laptop Ultra on NVIDIA RTX Spark; Microsoft Execution Containers Now GA; DGX Station for Windows; ASUS, Dell, HP, Lenovo, and MSI Systems Ship October 16
https://blogs.windows.com/windowsexperience/2026/10/07/building-windows-for-hybrid-intelligence/

Marvell Investor Day Sets $70B to 90B Fiscal 2031 Revenue Target; Fiscal 2028 Outlook Raised to~ 20B; XPU Attach Connectivity Leads Across Scale-In, Scale-Up, Scale-Out, and Scale-Across
https://x.com/PatrickMoorhead/status/2107480063381745966
https://www.cnbc.com/investingclub/2026/10/06/what-marvells-rosy-long-term-guidance-means-for-our-ai-chip-stocks.html

Anthropic Launches Claude Haiku 5.5, Third Model in the Claude 5.5 Family, at ~75% Lower Average Task Cost Than Haiku 4.5
https://www.reuters.com/business/anthropic-launches-third-claude-55-model-expanding-ai-lineup-before-planned-ipo-2026-10-07/

Mistral Launches Mistral Large 4
https://mistral.ai/news/mistral-large-4/

Reflection Unveils Beam, a 501B-Parameter Open-Weight Model With 23B Active Parameters
https://reflection.ai/blog/introducing-beam

OpenAI, Anthropic, Google, and Meta Testify Under Oath Before the New York City Council on AI Risks, Liability, and Shutdown Capability
https://www.cbsnews.com/newyork/news/new-york-city-council-ai-oversight-hearing/

Google Cloud Unveils the Gemini Agent at Gemini at Work 2026; Persistent Context Across Workspace, Microsoft 365, and Slack; Routes Work Across Gemini and Claude Models
https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026

The Flip

Are temporary construction jobs a strong reason to approve AI data centers?

FOR: Data center construction is the largest apprenticeship engine in decades, and the skills and paychecks outlast every project.
https://www.cnbc.com/2026/09/26/blue-collar-jobs-ai-data-center-backlash.html

AGAINST: Peak construction headcount rarely becomes durable employment, and jobs promises should become measurable obligations before approval.
https://www.cnbc.com/2026/10/03/data-center-backlash-europe-asia-africa.html

Bulls & Bears

Chip Stocks Lead Thursday’s Pullback as Software Holds Up Better
https://finance.yahoo.com/news/intel-slides-3-chip-stocks-140543692.html

NVIDIA (NVDA) Hits All-Time High Above $240, Nears $6 Trillion Market Cap, Then Gives Back the Milestone
https://x.com/danielnewmanUV/status/2107480808655294965

Samsung Electronics Q3 2026 Earnings Guidance; ~107.4 Trillion Won Operating Profit (~$80B) on ~195 Trillion Won Sales; First Tech Company Above 100 Trillion Won in a Quarter
https://news.samsung.com/global/samsung-electronics-announces-earnings-guidance-for-third-quarter-2026