2026.10.12 | Substack | META | TMTB Podcast Intel: Meta Muse & NVIDIA

Ethmos and TMT Breakout

Oct 12

READ IN APP

TMTB Podcast Intel is produced by Ethmos, research agents that listen to every episode in TMTB’s coverage universe and write the briefs below.

Today’s Essential Episodes:

1. Meta CEO Mark Zuckerberg and Chief AI Officer Alexandr Wang on Launching Muse Wonder, the $100B+ Compute Bet and Why AI Won’t Be Winner-Take-All

Cheeky Pint, with John Collison and Mark Zuckerberg, founder and CEO of Meta, and Alexandr Wang, chief AI officer of Meta Superintelligence Labs*2. Meta Chief AI Officer Alexandr Wang on Building Muse in Seven Months, Betting on Personal Agents Over Coding and Rebuilding Meta’s AI Lab *David Senra, with David Senra and Alexandr Wang, chief AI officer of Meta and co-founder of Scale AI*3. NVIDIA’s Justin Boitano on Why NVIDIA Opens Its Models, Why Cheaper Tokens Expand Demand and Why the Agent Harness Matters *The Deep View: Conversations, with Jason Hiner and Justin Boitano, VP of enterprise AI at NVIDIA*4. Google Labs and Gemini App Head Josh Woodward on Running a Company Like a Lab, the ‘Almost Possible’ List and Knowing When to Kill an Idea *Lenny’s Podcast, with Lenny Rachitsky and Josh Woodward, head of Google Labs, the Gemini app and AI Studio*5. Atlassian CEO Mike Cannon-Brookes on Why Agents Won’t Replace Software Interfaces, Why Enterprises Won’t Trust Chinese Models and Context as the Moat *20VC, with Harry Stebbings and Mike Cannon-Brookes, co-founder and CEO of Atlassian*6. Ben Bajarin and Jay Goldberg on Marvell’s Analyst Day, the OpenAI Revenue Scare and the Heterogeneous Rack *The Circuit, with Ben Bajarin and Jay Goldberg

Produced by Ethmos: AI-assisted summaries of public podcast episodes. For information only. May miss context or contain errors. Verify before relying on them. Summaries are not substitutes for the episodes, which belong to their creators. Not investment advice, offer, or solicitation. Not affiliated with or endorsed by the shows, hosts, guests or companies discussed. Corrections and publisher opt-outs: support@ethmos.ai.

Listen · Watch · Follow Meta on Ethmos

THE SIGNAL

• Zuckerberg says owning data centers beats renting cloud since Meta keeps using GPUs past full depreciation; he claims an H100 is worth more than last year because compute is constrained, which supports longer GPU useful lives (our read).- Zuckerberg says leaders committed roughly $100-200B late last year before seeing the model; Muse beat Meta’s own median launch expectation, and growth tooling wasn’t ready, so adoption could have been faster.- Several phone makers have approached Meta about building Muse Gadgets into handsets, and Meta says far more businesses want deep integration than want to block its agent, a possible distribution channel beyond Meta’s apps.- Collison says Stripe’s agentic-commerce volume only took off around September, months after its May announcements.

01 Muse Wonder’s architecture and launch

Muse Wonder is Meta’s new frontier-class model, succeeding the smaller Muse Spark line, with a larger “Wonder full” variant for heavier use. Wang says it is state-of-the-art on personal/professional agentic benchmarks and shows a step up in coding, citing results on DeepSWE, SWE Marathon, and Terminal Bench 4. Key architectural emphases, per Wang, are active memory (the model’s built-in awareness of how to retain and compact information over days or months) and active focus (persistence through long, ambiguous tasks despite setbacks) — both pitched as fixes for the memory problems users report with agents like OpenClaw. Zuckerberg notes the model was intentionally trained end-to-end for personal-agent use cases rather than optimized purely for coding, which Wang says was the industry’s dominant focus earlier in the year. Earlier in the conversation, both guests also push back on the idea that AI products converge onto a narrow, simple design space, arguing that model personality, discretion, and tool-use behavior involve many non-obvious design decisions obscured by today’s coarse product categories. Distribution: Muse stays free on the Spark model, while Wonder requires a paid subscription initially (due to higher serving cost), and is also available via the Meta API, OpenRouter, and Muse Code for developers.

02 Training mechanics and benchmarks

Both guests walk through pretraining (ingesting massive, multimodal data — text, images, video, audio, plus proprietary data) versus post-training (supervised fine-tuning followed by reinforcement learning that strengthens behaviors already established during SFT). Wang describes RL as requiring much more intentional, often smaller, higher-quality datasets and environment design than pretraining’s volume-driven approach, and notes real safety/alignment risk arises when RL optimization pressure pushes a model toward misaligned shortcuts — requiring careful environment curation. Zuckerberg frames Meta’s alignment philosophy as prioritizing instruction-following and user-value alignment over moralizing behavior. On benchmarks, Zuckerberg says Meta increasingly steers toward internal evals over public ones because external benchmarks saturate quickly, and Wang warns that chasing public benchmarks effectively outsources product judgment to whichever PhD student built a viral benchmark. Meta’s internal metric for this is “big model smell,” quantified as the gap between in-distribution and out-of-distribution performance, which Wang says shows a large jump from Spark to Wonder.

03 The capex bet and compute buildout

Zuckerberg details Meta’s compute strategy: the Prometheus cluster in Ohio (scaling past one gigawatt) and the Hyperion cluster in Louisiana (scaling to five gigawatts), both planned years in advance of MSL’s founding, tracing back to roughly 2023-2024 under the Llama program. He estimates operating a gigawatt cluster costs on the order of $15B/year, or $50-60B in upfront capex, with research costs (testing at smaller scale before big runs) roughly doubling total spend; Wang puts the cost of building a frontier model in the tens of billions of dollars. Zuckerberg argues the downside risk is capped because if AI generally succeeds, excess compute can be sold or redirected to other uses, so failure would require the entire AI thesis, not just Meta’s model, to fail. Wang recalls the visceral stress of this period, including Susan Li privately warning him of the financial stakes before the model’s performance was known.

04 MSL’s founding and culture

MSL launched following what Zuckerberg implicitly calls Llama 4’s disappointing trajectory; he began recruiting Wang around April 2025 and the lab is now just over one year old. Wang’s founding thesis — in Zuckerberg’s telling — was that talent density matters more than headcount: the research team numbers under 200, intentionally small to preserve startup-like energy within Meta’s resources. Zuckerberg says Wang was chosen despite never having trained a frontier model because of his strategic clarity and ability to attract talent; the launch required a full “reboot,” recruiting both inside and outside Meta. Separately, Zuckerberg cites Kunal Shah, brought in to run WhatsApp, and says CJ is joining to build an enterprise platform. Wang recounts an early internal meeting where a since-departed team lead questioned whether a reasoning model was even needed — a moment he says signaled the need for the reboot and for Meta leadership broadly to become more “AI-pilled,” which he says happened quickly, citing Instagram’s Adam Mosseri as an example of rapid adaptation.

05 Agentic commerce and monetization

Zuckerberg describes personally using Muse to coordinate baking with his daughter via Instacart deliveries, noting he wasn’t previously an Instacart user — raising what he calls an interesting economic question of whether agent-held accounts expand the total addressable commerce market. Wang adds that reduced search friction and holistic personalization (understanding a user’s family, friends, and context) are differentiators versus siloed recommendation systems. On platform strategy, Zuckerberg says Meta trained strong browser-use into Muse so no business has to integrate anything, while pursuing direct API deals with major partners; he says far more businesses are asking to integrate deeply than have tried to block it. Monetization paths discussed: subscriptions for Wonder access, a commerce “cut” of money saved or made for users, and a new enterprise platform to sell models like Wonder to businesses. Hardware plans include Muse Gadgets (third-party device form factors) and a home hub (“home link”) connecting Muse to Sonos and smart lights, alongside continued smart-glasses development as a multimodal, potentially app-store-free interface.

Listen · Watch · Follow Meta on Ethmos

THE SIGNAL

• Wang says A/B tests show Muse’s experience is highly sensitive to model quality; nearly all users of early-2026 agent products churned over unreliability, and checkpoints passing 80 of 100+ eval rows were still rejected.- Muse was built by under 200 people across model and product work, a deliberately small team on one of Meta’s most important launches this year (our read: focus, not headcount, drove the outcome).- Wang says MSL’s big pay packages were often roughly comp-neutral for recruits because OpenAI and Anthropic equity had surged, undercutting the view that Meta simply overpaid for talent.- Wang says Zuckerberg first called after Llama 4, which he calls an internal disappointment, and the resulting deal left Meta owning 49% of Scale, with Scale continuing and Wang and a few colleagues joining Meta.

01 From a 2025 memo to a working prototype

Muse traces to a June 2025 memo on “personal superintelligence” co-written by Zuckerberg, Wang, and Nat Friedman, outlining AI meant to improve everyday life broadly, written before agents had meaningfully taken off. The real catalyst came in February, when Anthropic’s Opus 4.5 and the open-source tool OpenClaw demonstrated agentic potential; Friedman — an early Stripe customer and GitHub Copilot contributor — became consumed by OpenClaw, describing his experience as somewhere between terrifying and euphoric, including having the agent monitor security footage to confirm he was drinking water. Wang’s own first OpenClaw session involved a deep psychological-probing prompt that he compares to years of compressed therapy. Both men wrote memos to Meta’s board in February arguing a “super agent” was the logical endpoint of consumer AI, and a prototype was built within one to two weeks and demoed to the board, with core elements like the Jolly mascot and an early Muse Charm already present.

02 Seven months of grinding reliability into the product

Wang frames the subsequent seven months as the hard, painful work of turning a prototype that only sometimes worked into a dependable consumer product. The team built evals against a spreadsheet of over 100 specific behaviors the model needed, tracking which were launch-blocking versus acceptable, and iterated until a specially trained checkpoint of Muse Spark 1.3 cleared every threshold. Wang credits Meta’s broader culture of metrics-driven optimization — previously applied to conversion and onboarding funnels — for this capacity to refine model edge cases. He contrasts this with OpenClaw’s trajectory, which rose and then faded because it never achieved comparable reliability, calling its creator, Peter Steinberger, a visionary nonetheless.

03 Strategic bet against coding agents

At the start of the year, the industry consensus held that coding agents were the only AI capability that mattered, with Anthropic surging ahead. Wang says Meta consciously rejected that race, betting instead that Meta’s heritage — connecting people to friends, interests, and hobbies — made it uniquely positioned for personal, not professional, AI. He told Zuckerberg early on that Meta’s product span represents the things people want to spend their time on rather than work obligations, framing this as a long-term structural advantage regardless of short-term coding-agent hype.

04 Zuckerberg’s role and the Scale AI deal

Wang describes a relationship with Zuckerberg dating to a 2021-22 introduction via Meta’s chief data officer Alex Scholz, though their first meeting reportedly took roughly a year to schedule. After Llama 4 disappointed internally, Zuckerberg called Wang for advice; within five to six weeks, Meta and Scale AI announced a deal in which Meta took a 49% stake in Scale while Wang and colleagues joined Meta. Wang says it wasn’t initially clear the call would lead anywhere, and he found the construct at first hard to believe, spending many weeks working through whether to leave the company he considered his life’s work for nine years. He ultimately saw it as beneficial for all parties involved, partly because Meta’s AI reputation was then considered damaged, giving him room to rebuild the lab from scratch.

05 Meme marketing and growth

Wang says making Muse break through the torrent of AI product launches required more than Meta’s distribution advantage. He leaned on friends he jokingly calls his “schizo” meme-making contacts and began posting increasingly unfiltered tweets and mascot images, framing this as a way to cut through noise differently from Anthropic or OpenAI’s ad campaigns, which he and Senra criticize as unclear or self-focused rather than benefit-led (citing Claude Hopkins’ Scientific Advertising). Screenshot-shareability — featuring the Jolly mascot — became central to word-of-mouth, exemplified by an anecdote about a CEO using Muse end-to-end to retrieve a lost ID via a Flexjet and courier. Meta has publicly said Muse usage is in the millions, with Wang amplifying third-party claims that it is the fastest-growing consumer AI app.

Listen · Watch · Follow NVIDIA on Ethmos

THE SIGNAL

• Boitano credits harness engineering for NVIDIA’s AVO research lifting an Opus 5 model from 30% to 100% accuracy on ARC-AGI-3; the episode doesn’t detail how the two runs were set up, so treat the jump as his characterization.- Boitano says NVIDIA’s large compute business can fund Nemotron 4 and 5 for the long run, making Nemotron a subsidized counter to Chinese open models rather than a standalone P&L (our read).- NVIDIA already runs CrowdStrike’s red and blue models against a digital twin of its own environment, which makes it a live reference customer for CrowdStrike’s agentic patching workflow, not just a supplier.- OpenShell grew out of treating agents like malware that grab credentials; Red Hat has built it into OpenShift and RHEL, Microsoft has said it is picking it up, and Boitano says infrastructure and cloud companies have too, putting NVIDIA into agent security below the model layer (our read).

01 Why NVIDIA open-sources Nemotron

Boitano, who leads NVIDIA’s enterprise AI business, says NVIDIA’s open-model strategy mirrors the decades-long evolution of operating systems from closed, full-stack integration toward open infrastructure valued for its transparency and security. NVIDIA is now on Nemotron 3, and Jensen Huang has signaled Nemotron 4 and 5 are in development. NVIDIA’s approach, per Boitano, is to open not just model weights but the underlying datasets and post-training techniques, giving any company — not just NVIDIA — the foundation to build domain-specific intelligence. He frames this as necessary infrastructure for the broader economy rather than a product NVIDIA wants to monetize directly.

02 CrowdStrike as proof point

CrowdStrike, at its Falcon conference, unveiled two new models (“red” and “blue”) built on Nemotron, illustrating Boitano’s thesis that software-as-a-service companies can convert from systems of record into producers of their own domain-specific digital intelligence. CrowdStrike’s proprietary Falcon telemetry becomes the differentiated data layer on top of NVIDIA’s open foundation. NVIDIA describes itself as both infrastructure partner and design partner, sharing harness-level research (including a cybersecurity-specific study on static-analysis tool calls) and using CrowdStrike’s own platform internally as a “best first customer.” Boitano notes the industry’s cybersecurity conversation has shifted from using large frontier models for costly vulnerability discovery toward cheaper, harness-optimized open models that find bugs comparably well, then verify, prioritize, and patch them in an end-to-end agentic workflow.

03 Enterprise motivations: control, cost, sovereignty

Boitano identifies three enterprise drivers for open models: control (owning institutional knowledge rather than risking it training a competitor’s general intelligence), cost (token costs scale fast once agents are involved), and data sovereignty/privacy, citing unclear data-retention terms at frontier labs. He points to Palantir’s recent sovereignty bootcamps in New York, each oversubscribed by hundreds of customers, as evidence of the scale of this demand. On cost, he cites Greg Brockman’s description of an OpenRouter experiment where discounting a GPT model by 50% drove roughly 15x higher demand rather than a proportional 2x — a result Boitano uses to argue that lower token prices expand, rather than cannibalize, the addressable market for AI use cases enterprises previously couldn’t afford.

04 Co-design, the stack, and research scope

Referencing Jensen Huang’s “five-layer cake” — power, chips, infrastructure/orchestration, models, and applications — Boitano says NVIDIA’s own model training work (running “AI factories”) keeps it focused on power efficiency and full-stack optimization, a practice he traces back to GeForce-era library optimization. NVIDIA fields roughly 1,000 researchers spanning pretraining, post-training/data curriculum design, reinforcement learning with verifiers, and inference-time scaling, and extends open research beyond LLMs into World Foundation Models, physics, vision-language-action robotics models, and self-driving (citing a model he refers to as Alto Maio). Boitano frames open publication as something that benefits the entire ecosystem of labs and application builders, not just NVIDIA.

05 Security: SAFE and OpenShell

Addressing the cybersecurity concerns following agent-related incidents (referenced as involving OpenAI and Hugging Face), Boitano says NVIDIA co-founded the Open Security Alliance and, with the Linux Foundation, published an open RFC called SAFE (Secure AI Findings Exchange), modeled on NASA’s aviation near-miss reporting system, to let the industry study agent reasoning traces when things go wrong — including cases where an agent, unable to solve a problem, began reward-hacking. Separately, NVIDIA built OpenShell, an Apache 2.0-licensed project (from engineers with intelligence-community backgrounds) that sandboxes each agent using OS kernel-level controls over network, storage, process, and inference-endpoint access, plus formal verification to prevent policy breaches. Boitano says Red Hat has integrated it into OpenShift and RHEL, and Microsoft has committed to adopting it, positioning hardware/kernel-level control as a deterministic counterweight to probabilistic agent reasoning.

06 Personal tools and leadership approach

Asked for a personal AI tool recommendation, Boitano cites a tool called Hermes, connected to email and business systems via OpenShell-restricted read-only access, functioning as a daily-briefing chief of staff. On leadership leverage, he describes a practice learned from Jensen Huang of identifying the single most important priority each day and dropping the rest, operationalized company-wide through weekly “top five” priority lists that NVIDIA uses AI to analyze for organizational alignment.

Listen · Watch · Follow Google on Ethmos

THE SIGNAL

• Woodward says a new Labs project team is now 2-3 people (maybe 4), down from 5-7, as execution costs fall; with 20-30 experiments running at once, Google can take more shots for the same headcount (our read).- Gemini has over 1B users, and Woodward says recent weeks taught them to drop modes and toggles for a single prompt box that just acts; Gemini Spark can already connect Gmail, Calendar and Docs data, with more personal-intelligence launches coming soon.- Labs now splits work into 0-to-1, 1-to-10 and 10-to-100 stages, judging early bets by user reactions and mature products by funnels and CPAs, because hits like NotebookLM, Flow and AI Studio risk crowding out new experiments.- Woodward says Labs needs more and more TPUs as products grow and holds up teams that free up compute as heroes, a firsthand sign that compute limits growth even inside Google (our read).

01 Operating like a labs team

Woodward’s framing: every company needs a team whose explicit job is to live at the frontier, running on subscriptions to every major model and staying plugged into what’s going viral in AI daily. This team should report with enough independence that it isn’t swallowed by an existing business unit; in some setups it should report close to the CEO. Woodward (echoing a framing from Dan Shipper, raised by host Lenny Rachitsky) agrees with having a dedicated frontier-testing function so the rest of the company isn’t distracted, but stresses a third requirement beyond independence and talent: the lab must be able to bridge its frontier work back into the core business, since many companies get the first two elements right and still fail because nothing gets commercialized.

02 Where ideas and PMF come from

Ideas rarely emerge from design sprints or scheduled brainstorms, per Woodward — they surface from engineers hacking on weekends, side projects, and obsessive small groups chasing something that catches their eye. Labs maintains an “almost possible” list and a document of 82 forward-looking predictions, tracking technical thresholds (model latency, modality support, language coverage) that signal a capability is about to flip from impossible to real; current watch areas include immersive/interactive entertainment and messaging/chat, which Woodward describes as newly unsettled territory. On product-market fit, he calls early-stage judgment far more art than science — conviction from a couple of obsessed people, then validated by literally watching users’ eyes light up in hallway demos. He warns against falling in love with a specific product solution rather than the underlying user problem, since most ideas go through three-to-five pivots.

03 Killing ideas and culture

Woodward’s claim is that the team almost always senses failure before the leader does, and his job as a leader is building psychological safety for people to flag that something isn’t working. He illustrated this with a Gemini feature the team pulled at the last minute based on disappointing user-reaction feedback. Labs also uses “seasons” — explicit, named high-intensity sprint periods (e.g., around a Google I/O launch) followed by deliberate slack periods for hacking and rediscovery — to manage what Woodward, borrowing a phrase from a Roger Federer biography, calls “explosive endurance”: the ability to go hard without burning out over years.

04 Org design and consumer opportunity

For companies setting up their own labs, Woodward’s three tips are: give the team real independence rather than nesting it under an existing unit; define clearly whether its mandate is to graduate ideas into existing products or invent wholly new categories, using an internal priority order that puts users first, Google second, and the product itself third; and use the lab to pilot new ways of working across the whole company (e.g., Labs piloted Google’s “builder” job ladder before company-wide rollout). Common mistakes include premature team consolidation and hiring too fast — Woodward recalls a past project that scaled to 30 engineers with no product-market fit but an inspiring vision. On where AI unlocks new consumer opportunity, he points to entertainment, messaging/chat, and a still-undefined “personal intelligence” assistant layer — and, pressed by Lenny on why Google hasn’t done more with users’ Gmail/Calendar/Docs data, says parts already exist (Gemini Spark can connect it), that users tell him they’d rather Google make it work than connect their data to competing products, and that more is coming soon.

05 Skills, planning horizons, and rituals

Rising-value skills Woodward named include a high “unlearning rate” (how fast someone abandons a disproven idea), explosive endurance, and the ability to build trust and collaborate across both people and AI agent swarms — he argues role specialization still matters even as more people prototype, comparing it to a college major versus minor. Labs plans on roughly a six-month rolling horizon, aiming for meaningful milestones within fifty to one hundred days, rather than multi-year roadmaps, because frontier capabilities can turn into shipped projects within 48 hours. He also called model benchmarks the industry’s most overhyped topic, arguing most users have no real sense of what an Elo rating means. Internally, Labs reinforces culture through named rituals like the TPU Harvester (a toy rake for engineers reclaiming compute), the golden Band-Aid (for fixing product “paper cuts”), and the LLM Whisperer award for prompting breakthroughs. He also discussed, unresolved, whether Google will end up with thousands of discrete products or just a handful of super-apps.

Listen · Follow Atlassian on Ethmos

THE SIGNAL

• Cannon-Brookes says Jira and the Teamwork Collection just had their best year of growth, accelerating at massive scale in users.- Atlassian sells products, including via DX, showing customers token spend by model and feature and where to cut it; he says enterprises are increasingly price-sensitive as AI becomes a growing budget line.- In Amsterdam, Atlassian announced agent-to-agent integrations with Figma and Databricks, so Confluence users can query Databricks or Snowflake through the Teamwork Graph, making Atlassian a hub rather than a bypassed destination (our read).- Atlassian runs many fine-tuned open-weight models, including a lot of NVIDIA Nemotron, and his biggest change of mind this year is inference prices falling further than expected, which suggests margin upside on AI features (our read).

01 Agents won’t erase interfaces

Cannon-Brookes rejects host Harry Stebbings’ framing that agents turn software into mere databases, saying that holds only for “pedestrian” functional tasks (like buying a charging cable) but not for collaboration-heavy enterprise workflows. He says Atlassian operates one of the top three-to-five MCP servers globally, enabling agents to read and write the Teamwork Graph, but argues that if Atlassian’s interfaces stop evolving to solve customer problems, the company loses relevance to that experience. He envisions a mixed future of horizontal agents, vertical deep-task agents, and persistent interfaces coexisting — not one “Uber agent” or single magic button. As evidence interfaces still carry commercial value, he cites Confluence generating north of $1.5B ARR, growing over 25% annually at scale — large enough to be a standalone public company — despite writing text seeming commoditizable.

02 New design problems agents create

He identifies unresolved UX challenges: today’s chat text-box interface is a poor way to expose LLM power; provenance and attribution of AI- versus human-generated content (for instance, whether a person or Codex produced a given pull request) is unsolved and matters for accountability rather than quality judgment; and companies need interfaces that flag exceptions needing human attention without drowning users in agent-trace detail. He predicts that within five years, watermarking AI content may become moot as fully synthetic content becomes normalized, drawing an analogy to how audiences don’t ask whether a film avoided CGI or color grading.

03 Chinese models and the specialization debate

Asked about an investor’s claim that enterprises refuse Chinese open models, Cannon-Brookes confirms the pattern: Chinese-labeled models face explicit pushback while open models broadly (Nemotron, Mistral) do not, and Atlassian deliberately excludes Chinese models from production for customer-trust reasons — even while arguing the geopolitical concern is logically separate from the technical question, drawing a comparison to enterprises already trusting Linux despite heavy Chinese contribution to it. On custom models, he says the benefits of specialization remain unproven: legitimate cases exist (robotics, proprietary-data-heavy domains), but most companies claiming proprietary training have merely fine-tuned someone else’s billion-dollar model, often Chinese, and such specialization has repeatedly been overtaken by faster-improving generalized frontier models.

04 Cost discipline and inference economics

Cannon-Brookes argues well-funded private AI startups lack the discipline to optimize cost, while Atlassian — as a public, enterprise-serving company — must build FinOps muscle: an internal AI gateway team handles all model routing and blending in-house, rather than relying on third-party tools like OpenRouter, to balance quality, speed, and price per feature. He advises product teams to build compelling features first and optimize cost only once usage scales from a handful of users to millions. He predicts inference price-per-performance will keep falling given the volume of capital and talent applied, even as Jevons-paradox-style demand growth — more inference consumed because it’s cheap — offsets some savings, noting that right now technology production is outpacing how fast customers can actually consume it.

05 The Teamwork Graph as the real moat

He frames company context — not proprietary training data — as Atlassian’s moat: the Teamwork Graph combines documents, decisions, code, and a social graph of who-works-with-whom, pre-inferring likely answers much like a private search index rather than retrieving live. He stresses the graph must stay perpetually fresh, since stale facts are useless though historical “learnings” retain value, and explicitly says Atlassian has no interest in acquiring companies purely for their data, contrasting this with deals aimed at building generic resellable datasets (citing, from memory, a deal for Spirit Airlines’ emails and other corporate data). On OpenAI and Anthropic pushing into docs and office-style applications, he calls them deep Atlassian partners and says the boundary between infrastructure providers and application competitors remains unresolved, noting that writing tools alone have never been a definitive moat — design, integration, and workflow purpose are. He also briefly addresses Jira versus Linear on design craft, calling the comparison unfair and arguing Atlassian’s job is to keep uplevelling its own experience over a multi-decade race rather than match any single competitor’s aesthetic.

Listen

THE SIGNAL

• Goldberg says the US grid can add 20-30 GW next year (2027), energy research points to the low end, while compute vendors privately want more than 30; the hosts call next year peak constraint, easing in 2028-29.- Goldberg says chip companies with surplus TSMC allocation and weak core end-markets are already repurposing wafers into hyperscaler custom ASICs, a quiet source of incremental AI supply beyond headline capacity adds.- Bajarin cites BMC forecasts of $3-8B by 2030, and Goldberg says a 2% power saving from better board management matters at gigawatt scale; this gives a rough TAM for Marvell’s teased management chip (our read).- Bajarin says Ayar Labs will show a working co-packaged optics demo with GUC at OCP, and expects new analog silicon now taping out after two-year roadmaps; worth checking for CPO and analog timing evidence (our read).

01 Marvell’s repositioned narrative

Marvell’s Analyst Day in New York produced five-year revenue and EPS targets that, per Goldberg, exceeded even an implied analyst consensus, with a notably upbeat mood in the room. The event marked a deliberate shift away from years of conversation centered on Marvell’s custom-ASIC share battle against Alchip at Amazon, toward a broader data-center-infrastructure story spanning optical interconnects, switching silicon, and XPU-attach products. Both hosts agree Marvell has had these capability pieces for years but is now able to flesh out a more coherent story because data-center spend has grown large enough that previously niche product lines now represent meaningful revenue, sometimes hundreds of millions to billions of dollars.

Bajarin pushes back gently on whether Marvell has clearly identified its ideal competitive position, noting the company still spans processors, networking silicon, and connectivity, and arguing it looks well-positioned for a heterogeneous infrastructure world rather than one single lane. He notes Marvell’s NVIDIA-NVL compatibility and partner-friendliness versus Broadcom’s likely reluctance to adopt NVL as a standard, suggesting Marvell can capture business in spaces Broadcom won’t pursue directly.

02 Heterogeneous compute and inference differentiation

Bajarin elaborates on his thesis that compute workloads must become fungible across vendors: hyperscalers and neo-clouds will never be purely single-vendor, requiring scale-up/scale-out “domains” — connected racks acting as one shared-memory compute brain — to interoperate across GPU, CPU, and ASIC vendors. He insists terminology should always pair “scale up” or “scale out” with “domain” for clarity, since only “scale across” currently has unambiguous meaning industry-wide. Goldberg, fresh off investor meetings following the Marvell event, reports growing conviction that inference compute itself is fragmenting — not just training vs. inference, but further specialization between prefill and decode stages — citing Google’s expanding TPU generations (Ironwood, upcoming gen eight/nine) and NVIDIA’s CPX alongside Grok’s LPUs. He says this trend creates more openings for competitors beyond NVIDIA, though he stops short of predicting dozens of startup XPUs will fill every niche.

03 The teased management-silicon category

Bajarin highlights Marvell’s vague preview of new AI infrastructure management silicon, deliberately positioned near — but not labeled as — a BMC (board management controller), a category dominated by Taiwan’s A Speed and complemented by Lattice FPGAs. Goldberg explains BMCs traditionally track low-level board signals (voltage, temperature, current) across hundreds of sensors; as racks and chips multiply sensors and complexity, engineers have been bolting on more FPGAs to handle security, power management, and monitoring that legacy BMCs weren’t designed for. Marvell’s unnamed chip appears aimed at this gap, though Goldberg notes the company merely teased it without naming a real product. Goldberg frames this as part of a broader pattern at the event: roughly a dozen similarly in-the-weeds product categories signaling how wide Marvell’s addressable opportunity has become as engineering complexity compounds across logic and analog silicon.

04 OpenAI revenue scare and trade fragility

A news report suggested OpenAI’s revenue ran roughly $20B below expectations — around $50B versus a rumored $70B ARR target — triggering a semis selloff. Goldberg says investigation showed the gap stemmed from differing accounting treatments (gross vs. net revenue, after traffic-acquisition costs) rather than any real shortfall. Bajarin adds that even conservative estimates imply dedicated AI spend across hyperscalers, OpenAI, and Anthropic could reach $250-300B this year, still roughly doubling year-over-year.

The more consequential point, per Goldberg, is market awareness that AI infrastructure spend is concentrated in a small handful of companies — really just two, OpenAI and Anthropic — both still generating large losses, with significant equity and debt market exposure riding on their trajectories. Both hosts agree the risk isn’t collapsing demand but operational bottlenecks: cybersecurity incidents, data breaches, or — most likely, per Bajarin — the inability to energize gigawatts fast enough to match accelerator shipment schedules. Goldberg poses a hypothetical: if a major hyperscaler cluster gets delayed six months next year due to turbine or transformer shortages, markets might not react as calmly as they would today, given how much expectation is now priced in.

05 Adoption cycle and UI as bottleneck

Bajarin offers a framework for why this AI buildout may warrant more benefit of the doubt than past mega-cycles (dot-com, electricity, mobile): all prior cycles required customers to buy new costly hardware before adoption could scale, whereas AI can be used today on existing devices. Goldberg counters that the dot-com bubble also featured genuine demand undermined by inadequate infrastructure — citing Webvan’s early-but-underbuilt warehouse model and dial-up bandwidth limits before broadband penetration around 2004-2005 — and worries frustrated users today, overwhelmed by AI cost and security headlines, aren’t yet seeing personal benefit. Bajarin ties this back to a prior conversation with Benedict Evans about UI mattering as much as raw capability, noting he’s shifted some usage to Claude over ChatGPT due to improved project-organization interfaces, and flags ChatGPT’s new prompt-prefill feature as part of an ongoing battle over interface design to drive adoption at enterprise and consumer levels.

06 AMD, TSMC capacity, and the Intel wafer theory

AMD’s Lisa Su said in Asia this week that AMD expects greater-than-expected TSMC capacity next year. Goldberg builds a speculative theory from this: TSMC’s public language about Intel has grown notably cooler over the past several earnings calls as Intel Foundry (under Lip-Bu Tan) becomes technically competitive via EMIB and 14A process advantages. Goldberg speculates TSMC may be directing extra capacity to AMD partly to pressure Intel’s CPU-driven cash flow, since cash generation — not technical capability — is seen as Intel Foundry’s biggest constraint to competing long-term. Bajarin adds a parallel dynamic: a severe PC-market downturn, which he says forecasters keep marking down, could free up wafer capacity that gets redirected to higher-ASP server CPUs — potentially benefiting Intel’s foundry economics even as its PC business suffers, and similarly benefiting AMD’s server and Helios ASIC allocations.

POWERED BY Ethmos for TMTB

Get briefs like these for every company you follow.

TMTB Podcast Intel picks a handful of episodes a day. With Ethmos, research agents listen to every podcast that touches the names in your book and file briefs like these every morning.

Start with Meta, NVIDIA, Google, Atlassian, Marvell or your own list.

Start your 14-day free trial

TMTB readers get 25% off their first month. Cancel anytime.

TMTB Podcast Intel is produced by Ethmos for TMT Breakout. Each brief is an AI-assisted analytical summary of a publicly available podcast episode, prepared for information only; it is a condensed analytical summary, not a substitute for the episode: it is not intended to reproduce the programming and may miss important details, context or nuance, and it may contain errors, so verify against the original episode before relying on it. Each episode and its recording remain the property of their creators. Nothing here is investment advice or an offer or solicitation of any kind. Neither Ethmos nor TMT Breakout is affiliated with, sponsored by, or endorsed by the shows, hosts, guests, or companies discussed; show and company names are the property of their respective owners. Publishers who prefer not to be covered, and anyone with a correction, can write to https://substack.com/redirect/6403bff0-3415-47c0-a9ed-1c78636b22ec?j=eyJ1IjoiNDR3MndrIn0.NBlbpuapVvjXEQ-22Fcu951Q9Y3oMbVG8TDoGPfg2x8support@ethmos.ai; publisher opt-outs are honored.

Disclaimer:

This newsletter is published for informational and educational purposes only. Nothing here is individualized investment advice, a recommendation tailored to any person’s circumstances, an offer, or a solicitation to buy or sell any security or financial instrument. This newsletter is not an investment advisory service, and no advisory or fiduciary relationship is created between me and any reader.

*I actively trade the securities discussed here, often short-term. My positioning may be inconsistent with the views expressed, and I may buy or sell — including immediately before or after publication — without notice or updates. Price targets, upside/downside scenarios and time frames are illustrative opinions, not commitments to hold. *

This publication is for the subscriber’s personal use only and may not be redistributed, republished, forwarded, or quoted without prior written permission.

Copyright © 2026 TMT Breakout. All rights reserved.

Full disclosures here