Meta and Google Rejoin the Frontier, Microsoft’s Maia 300, and the AI Credit Scare | The Six Five Pod Ep. 324

On Episode 324 of The Six Five Pod, Patrick Moorhead and Daniel Newman unpack the biggest stories from the week, including new agent models from Meta, Google, and NVIDIA, Microsoft's Maia 300, Applied Materials' EPIC Center, AMD, NVIDIA, and Cisco at OCP Global Summit 2026, and the Apple-Meta Muse Spark Open partnership, plus a Flip debate on whether AI leaderboards serve vendors more than enterprise buyers. Plus, Bulls & Bears covers the CPI and PPI shock that rattled the AI trade, Anthropic's investor event, ASML's record quarter, and TSMC's earnings results.

The handpicked topics for this week are:

1. Meta, Google and NVIDIA Advance Agent Models, With Reliability Still Uneven: Meta launched Muse Wonder on October 12, a closed frontier model available through its API and Muse Code at $4 per million input tokens and $20 per million output tokens. Dan says it performs well on coding but is expensive out of the gate. Google opened Gemini 4 Argon broadly on October 13, and Dan notes it took the top spot on the Signal65 Pinnacle composite leaderboard, ahead of GPT and Claude models. NVIDIA followed with Nemotron 3.5 Super VL, a 120-billion-parameter open model, and Nemotron 3.5 Bolt, which fits in 24GB of memory. Pat points out how far models have come: Llama 3 scores 269 on Pinnacle against roughly 24,000 for Argon. He adds that Nemotron earns a perfect openness score from Artificial Analysis and that Bolt runs on DGX Spark and RTX Spark laptops. Both agree Meta and Google, written off earlier this year, are back in the frontier race. (The Decode)

2. Microsoft's Maia 300 Makes Custom Silicon Central to Its Inference Strategy: Microsoft announced the Maia 300 on October 12, its third-generation inference accelerator, and said Maia will become the primary inference system for its own products as deployments scale to multi-gigawatt levels from 2027. Microsoft expects 40% more tokens per megawatt than leading GPU-based systems, which Pat reads as a Vera Rubin comparison and rejects the "end of NVIDIA" narrative, arguing Maia is about Microsoft's margins on its own SaaS and PaaS workloads rather than raw chip performance. He also credits Microsoft with reaching this point in three to four years, a timeline that took Google and AWS closer to a decade. Dan notes the system's memory capacity and calls the third generation the moment custom silicon typically lands. He adds that Microsoft is funding its buildout from operations, without equity raises or heavy debt, which positions its margins to improve as custom chips scale. (The Decode)

3. Applied Materials' EPIC Brings Chip R&D Teams Together to Shorten Development: Dan served as master of ceremonies for the October 12 unveiling of Applied Materials' EPIC Center in Sunnyvale, where founding partners TSMC, Intel, Samsung, Micron, and SK hynix will place engineers alongside Applied Materials' teams. Dan explains that the center targets parallel development across transistors, interconnects, memory, and advanced packaging, including backside power and hybrid bonding, while each partner protects its own IP. Pat says EPIC's value will ultimately be measured in time to qualified production and yield. He frames it as a bet on removing the sequential handoffs that have historically slowed equipment and process co-development. (The Decode)

4. OCP Global Summit 2026: AMD, NVIDIA and Cisco Shift the AI Race to Rack-Scale: Pat frames the rack as the smallest unit of AI compute and the data center as the largest. At the OCP Global Summit 2026, AMD introduced Helion, an agentic AI rack offered as a CPU-only configuration with up to 32,768 cores or an enterprise configuration pairing EPYC CPUs with Instinct MI350P GPUs, in air- and liquid-cooled versions. Lisa Su also showed Verano, the first EPYC with an LPDDR memory interface. NVIDIA expanded its MGX rack contributions and 800 VDC facility designs, and NVIDIA and Cisco proposed a denser, field-replaceable pluggable optics standard. Pat notes interface standards that once lasted five to seven years now turn over in about a year. Dan highlights the gap between market fears about AI and the pace of industry collaboration on display. (The Decode)

5. Apple and Meta Bring Large Open-Weight Agents to the Mac Desktop: Meta released Muse Spark Open on October 15, an 816-billion-parameter open-weight model with 42 billion active parameters, tuned with Apple to run locally on clustered M5 Ultra Mac Studios. Mark Zuckerberg credited Apple CEO John Ternus for the collaboration. Dan sees Meta gaining access to Apple developers while Apple gains frontier models it failed to build itself, and he questions whether Apple is settling into the device layer. Pat calls it a competitive response to NVIDIA's RTX Spark laptops and DGX Spark, landing a $40,000 to $50,000 price point that NVIDIA doesn't cleanly cover. (The Decode)

6. The Flip — Are AI leaderboard rankings becoming more useful to vendors than to enterprise buyers? Patrick takes the FOR side: suggesting leaderboards built on throughput, intelligence, and cost per token serve hyperscalers and neoclouds. Daniel takes the AGAINST side: arguing that vendors grading themselves is marketing, but independent rankings force vendors to publish charts where they lose. Leaderboards cut 50 models down to five, and scoped testing on real jobs, like Signal65's, tells a CIO which one survives before signing a PO. Off the flip, both agree cost per token alone has lost its meaning. (The Flip)

7. CPI/PPI Revive Rate-Hike Fears and Pressure the AI Trade: Dan traces the midweek capitulation to energy-driven inflation pushing CPI, PPI, and bond yields higher, raising the cost of debt for capital-intensive growth companies. CoreWeave became the canary in the coal mine with more than $40 billion in debt and bond yields, Dan says, spiked toward 16%. The pressure spread to neoclouds like Nebius and IREN, then to chipmakers and semiconductor equipment names, with some stocks falling 10% to 20% in a day. Dan calls the move overblown, pointing to CoreWeave's roughly $105 billion backlog and its anchor customers Microsoft and Meta. He also notes the selloff reversed sharply a day and a half later. (Bulls & Bears)

8. Anthropic's Investor Event Puts Growth and Adjusted Profitability Under Scrutiny: At its October 14 investor day, Anthropic reported third-quarter revenue of $16.8 billion, up 45% sequentially after more than doubling the prior quarter, and said it was profitable on an adjusted basis. Some investors have since discussed a valuation near $1.5 trillion rather than the roughly $2 trillion previously expected. Pat compares the math to Arm at about 49x trailing sales and Palantir above 40x, but notes Anthropic carries more variability. Dan hopes the book runners price conservatively so the IPO trades up, and warns that "adjusted" is doing a lot of the lifting. He adds that Fortune 100 CFOs he spoke with this week see enough token demand for the frontier labs and open models to keep growing. (Bulls & Bears)

9. ASML's Record Quarter Puts Capacity Execution Back in Focus: ASML beat and raised, signaling it needs to expand output to meet demand tied to a $10 trillion+ AI infrastructure forecast. Dan calls the result written in the stars and finds the market's muted reaction odd for a company with a monopoly on EUV lithography. Pat explains that ASML's pricing is locked into contracts far in advance, so its guide and machine capacity are more useful metrics than a beat. He also notes that about 10% of ASML's GAAP EPS came from the roughly half-billion-dollar increase in the value of its Mistral stake. (Bulls & Bears)

10. TSMC's Results Show Strong AI Demand, With Capacity Still the Constraint: TSMC beat on revenue and net income, with demand strong enough to raise its outlook, while capacity allocation, packaging, and ramp costs determine how quickly that demand converts to earnings. Pat flags that TSMC did not raise its capital spending for the year, and deferred questions about Elon Musk's TerraFab to its next investor conference. Dan notes TSMC's costs are less exposed than hyperscalers' to the runaway memory prices driving much of their CapEx growth. Pat highlights TSMC's message that it won't leave table scraps for competitors. (Bulls & Bears)

Intro

Elon Flying Car Clip
https://www.youtube.com/watch?v=BYWy6SyJRZo

The Decode

Meta, Google and NVIDIA Advance Agent Models, With Reliability Still Uneven
https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/

Microsoft’s Maia 300 Makes Custom Silicon Central to Its Inference Strategy
https://moorinsightsstrategy.com/field-notes/microsoft-commits-to-maia-at-gigawatt-scale-and-as-its-future-primary-inference-platform/

Applied Materials’ EPIC Brings Chip R&D Teams Together to Shorten Development
https://x.com/danielnewmanUV/status/2109781773546844497

OCP: AMD, NVIDIA and Cisco Shift the AI Race to Rack-Scale Systems
https://newsroom.amd.com/news/open-rack-blueprint-scale-agentic-ai/

Apple and Meta Bring Large Open-Weight Agents to the Mac Desktop
https://www.cnbc.com/amp/2026/10/15/apple-meta-partner-ai.html

The Flip

AI leaderboard rankings are becoming more useful to vendors than to enterprise buyers.

FOR: A leaderboard win is a marketing claim until it transfers to the buyer’s work.
https://x.com/alexandr_wang/status/2109707211249537348

AGAINST: Better, transparent evaluations are the antidote to marketing.
https://x.com/Signal_65/status/2110764509702828084

Bulls & Bears

CPI/PPI Revive Rate-Hike Fears and Pressure the AI Trade
https://www.reuters.com/business/sp-500-dow-futures-rise-ahead-big-bank-results-cpi-data-2026-10-14/

Anthropic’s Investor Event Puts Growth and Adjusted Profitability Under Scrutiny
https://x.com/PatrickMoorhead/status/2110839356722380954?s=20

ASML’s Record Quarter Puts Capacity Execution Back in Focus
https://x.com/danielnewmanUV/status/2110317150443331829?s=20

TSMC’s Results Show Strong AI Demand, With Capacity Still the Constraint
https://pr.tsmc.com/english/news/3346

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