• Consider NVIDIA (NVDA) for long-term AI exposure: the view is strongly bullish on sustained data-center demand and recurring chip upgrades, though the $10 trillion-in-five-years claim is speculative.
• TSMC (TSM) is the clearest direct buy idea: it was praised as an exceptional, attractively priced supplier to the AI-chip industry, with no specific price target given.
• For broader AI exposure, monitor power and data-center infrastructure; rising electricity demand may create opportunities, but no specific securities or timeframe were identified.

NVIDIA (NVDA)

• Walter Goodwin called NVIDIA an extraordinary business with leading AI chips, systems, and data-center inference capabilities.
• He said confidently that NVIDIA could become a $10 trillion company within five years. He argued that AI infrastructure spending may last longer than skeptics expect, and that chips will need replacing as data centers upgrade.
• At the same time, he sees opportunity for specialized chipmakers: certain workloads may benefit from chips that offer more speed or efficiency than general-purpose GPUs. He also argued that power and data-center capacity constraints make efficient chips more valuable.
• Goodwin said NVIDIA’s acquisition of Groq addressed a need for faster inference, but suggested SRAM-based chips such as Groq’s face limits in memory capacity, cost, and context length.

Takeaways

• The discussion presents a strongly bullish long-term view of NVIDIA, grounded in continued AI demand and recurring hardware upgrades—not a specific near-term price target.
• Treat the $10 trillion in five years statement as the CEO of a prospective competitor’s opinion, not a guaranteed outcome. Specialized chips and constraints on data-center power and capacity are relevant uncertainties.

TSMC (TSM)

• Goodwin singled out TSMC as an exceptional supplier underpinning much of the world’s frontier chip production.
• He said he would “buy that stock every day of the week,” describing the company as an extraordinary business and calling it “cheap at the price.”
• The conversation also highlighted industry dependence on limited manufacturing capacity, including TSMC process capacity and memory supply.

Takeaways

• The episode offers a directly bullish view of TSMC as a key supplier to the AI-chip ecosystem.
• The supply constraints discussed could support the strategic importance of chip manufacturing, but they also underscore how dependent chip companies are on access to production capacity.

Fractile (Private)

• Fractile is building specialized AI chips and rack-scale systems, with a focus on making inference faster and more efficient.
• Goodwin described its approach as stacking memory over processing logic to combine high memory bandwidth with greater capacity. He argued that this could help run long-context models and extended AI-agent workloads.
• The company had raised a $720 million Series C at a reported $7 billion valuation. Goodwin said its silicon had returned from the foundry, with further trials and full product tape-outs planned.
• He described the company’s in-house manufacturing expertise and direct supplier relationships as advantages. He also said the company intends to remain a chip supplier rather than compete with customers by selling its own models.
• Risks he identified include the difficulty of reaching volume production, securing supply-chain capacity, and getting the post-tape-out work right—including system reliability and manufacturing yields.

Takeaways

• Fractile represents a high-risk, high-potential private-market bet on specialized AI hardware and inference growth.
• The investment case depends on more than the chip design: the company still needs to demonstrate reliable systems and production at meaningful scale. The large funding round and valuation are not, by themselves, proof of commercial success.

AI Inference Providers (Private)

• Goodwin was broadly bullish on companies that host and serve AI models, naming Modal, Fireworks, Baseten, and Together. He declined to pick one.
• He said he would rather invest in an inference provider than a neocloud today, arguing that inference providers can build a business flywheel around enterprise customers consuming tokens.
• He suggested the sector could produce a couple of multi-hundred-billion-dollar companies, while cautioning that it is unlikely to support ten such leaders.
• He sees room for providers to specialize—for example, in multimodal models or fast inference—but said the market’s long-term structure is still uncertain.

Takeaways

• The discussion identifies inference services as a potentially attractive private investment theme, supported by expectations that model usage and inference spending will grow.
• Goodwin’s comments do not distinguish among the named providers. Investors would need to assess each company’s customer traction and ability to differentiate; the transcript provides no specific company-level recommendation.

Nebius (NBIS) and Neoclouds

• The episode discussed neoclouds—specialist cloud providers built around AI compute—and their efforts to expand into inference and model routing.
• Goodwin agreed that neoclouds are trying to move beyond basic compute as they face the risk of commoditization. He said “one neocloud that is competent is much like another” and described the space as one that could produce both significant winners and losers.
• Nebius (NBIS) was mentioned in connection with its interest in model routing and inference, but Goodwin did not make a specific investment call on the company.
• Fluidstack, a private company, was cited as an example of a neocloud that moved away from relying solely on NVIDIA GPUs by deploying Google TPUs. Goodwin said it had an extraordinary year after making that shift.

Takeaways

• Neoclouds may offer upside, but the episode’s tone is cautious about differentiation: basic compute could become a commodity, making vertical integration and access to alternative chips important.
• Nebius and Fluidstack are examples of strategic approaches, not endorsed stock picks. In particular, the transcript does not provide a valuation or specific outlook for Nebius.

Alphabet (GOOGL)

• Google’s TPUs were discussed as an alternative to NVIDIA GPUs. Goodwin pointed to Fluidstack’s deployment of TPUs as evidence that neoclouds may be able to diversify their hardware.
• The discussion framed access to different chip platforms as one way a cloud provider might improve its economics and reduce dependence on a single vendor.

Takeaways

• Google’s TPU ecosystem may benefit from the search for alternatives to NVIDIA GPUs, particularly among cloud and AI-compute providers.
• The episode provides a strategic demand signal, not a direct assessment of Alphabet’s stock or its TPU business economics.

Broadcom (AVGO)

• Goodwin described Broadcom as a backend partner for several major AI-chip efforts, including platforms associated with Google, Meta, OpenAI, and others.
• He argued that using a common backend and similar manufacturing building blocks can result in chip platforms that are less structurally differentiated. Fractile’s approach, by contrast, aims to design and manage more of that work in-house.

Takeaways

• Broadcom appears in the discussion as an important AI-chip supply-chain participant, but Goodwin did not offer a direct view on its stock.
• His comments point to a possible trade-off: shared backend capabilities can support chip development, while companies seeking more distinctive architectures may try to control more of the design and manufacturing process themselves.

Specialized AI Chipmakers (Private)

• The discussion referenced Cerebras, Groq, Etched, Fractile, and Dematrix as examples of non-mainstream or specialized chip efforts.
• Goodwin said specialized chips may deliver substantial advantages on selected workloads, but also highlighted constraints. In particular, he said SRAM-based chips can offer fast inference but are limited by memory capacity, cost, and the length of context they can handle.
• He expressed confidence in Dematrix, which he said is pursuing a different architectural approach in the same broad 3D-memory area as Fractile. He said he believed both companies could be successful.

Takeaways

• The theme is bullish on specialized AI hardware, especially where it can improve inference speed or efficiency versus general-purpose chips.
• These are early-stage, private companies, and the transcript highlights meaningful execution and production challenges. Goodwin’s positive view of Dematrix is an opinion, not evidence that its technology or commercial prospects are proven.

AI Energy and Data-Center Infrastructure

• Goodwin described land and power as major constraints on expanding AI compute. He said grid-connection queues in the US and UK indicate demand far beyond available grid capacity, while acknowledging that not every proposed project will necessarily proceed.
• He suggested that investors seeking broad exposure to AI could be “long AI by energy,” because power is a fundamental input to data-center growth.
• He argued that more compute will be needed but that growth may come from delivering more intelligence per megawatt, rather than building thousands of times more data centers.
• He identified renewables and nuclear power as potential contributors to future supply, and said China has an advantage in moving quickly on large energy projects. He also pointed to permitting and local opposition as obstacles in the US, UK, and Europe.

Takeaways

• The episode highlights energy supply, grid access, and data-center sites as critical infrastructure themes for AI investment.
• The potential opportunity spans power generation and data-center infrastructure, but the transcript also identifies permitting, grid connection, and project delivery as constraints. No specific energy company or security was recommended.