← All issues

Inference, AI Agents, and the New Power Stack

Issue

Inference, AI Agents, and the New Power Stack

Distilling Deep AI Insights from Leading Podcasts

3 episodes3 podcastsNovember 22, 2026

This episode focuses on the emerging economics of AI compute and why GPU infrastructure may become one of the most important markets in tech. Kush Bavaria argues that compute should be treated like a tradable commodity, with standardized pricing, indices, futures, and hedging products similar to oil or other major commodity markets. The discussion also covers why GPU prices remain elevated, why older GPUs still retain value, and how neocloud providers like CoreWeave and Nebius may benefit from persistent demand. Beyond infrastructure, the episode explores market structure, suggesting that AI may evolve toward a horizontal compute layer plus many specialized inference companies. It also touches on startup strategy, fundraising, culture, and the broader question of whether compute or talent will be the ultimate bottleneck in AI.

Key Takeaways

  • 01

    AI infrastructure providers may capture durable value because compute can behave like a rent-generating asset class.

    The argument is that GPU suppliers and data-center operators can continue earning recurring revenue as long as AI demand stays strong, even if individual model companies come and go. This makes the business model more resilient than many assume, especially if compute remains scarce and essential to training and inference. The concept is closely tied to neoclouds, GPUs, and data-center leasing economics.

  • 02

    Compute may evolve into a commodity market with prices, indices, and hedging tools.

    Kush frames compute as something that should trade more like oil than like a normal software input, because many buyers and sellers are dealing in largely standardized capacity. If that happens, financial tools such as futures, indices, and hedges could become critical for managing price volatility and supply risk. This would mark a major shift in how AI infrastructure is financed and traded.

  • 03

    GPU prices are not collapsing; demand is keeping newer chips expensive.

    A common assumption is that GPU prices should fall quickly as supply catches up, but the conversation argues the opposite is happening for leading-edge chips like Blackwell and Rubin-class hardware. Persistent demand from AI workloads is supporting elevated pricing, which changes assumptions about margins and depreciation. This is important for buyers, lenders, and investors trying to forecast returns on AI hardware.

  • 04

    Older GPUs still have meaningful residual value because inference demand extends their useful life.

    The episode challenges the idea that older chips become obsolete quickly, noting that they can still be rented out for inference and other less demanding workloads. That means depreciation may be slower than the market expects, which is good for asset owners and financing structures. This matters for pricing, underwriting, and secondhand GPU markets.

  • 05

    The AI market may separate into a horizontal compute layer and many specialized inference providers.

    Rather than one dominant all-purpose inference company, the speaker expects a more fragmented ecosystem where a base infrastructure provider supports a range of specialized products. That suggests competitive advantage will come from owning infrastructure and building targeted tools on top of it. The idea reflects a platform-plus-application model in AI infrastructure.

  • 06

    In AI, access to compute may matter as much as or more than model talent.

    While strong researchers and engineers still matter, the episode argues that compute can be the decisive constraint because more compute can directly improve model quality and output. In that view, companies with better access to chips and fabrication capacity may outcompete better-branded or better-funded rivals. The discussion references TSMC, GPU supply, and model training capacity as strategic bottlenecks.

This episode centers on Cosign’s approach to professional networking, which prioritizes reputation, trusted recommendations, and peer vouching over static resumes or traditional profile pages. The guests discuss how AI agents are becoming an important new interface for discovery, with users connecting agents to Cosign so they can continuously search for people, lists, and opportunities. A major theme is that the most valuable professionals are often the least visible publicly, especially operators and builders whose impact is hard to capture on LinkedIn. The conversation also covers the Cosign Awards, which are designed to surface talented people through curation rather than create a rigid ranking. Overall, the episode imagines a future where both humans and agents participate in a more dynamic, relationship-driven professional network.

Key Takeaways

  • 01

    Cosign is building professional discovery around reputation and trusted recommendations instead of static resumes or profiles.

    The core product idea is that what matters most is not just what someone says about themselves, but who can credibly vouch for their work. This is significant because platforms like LinkedIn often fail to capture real contribution, especially for people doing important behind-the-scenes work. By shifting toward reputation signals, Cosign aims to make professional discovery more accurate and more useful.

  • 02

    AI agents are becoming active users of professional networks, not just human searchers.

    A notable emerging behavior is people connecting agents to Cosign through MCP so they can continuously look for relevant people, lists, and opportunities. This matters because it turns networking from a one-time search into an always-on discovery layer. The concept suggests a future where tools like MCP-enabled agents can surface opportunities proactively rather than waiting for a manual query.

  • 03

    The platform is especially valuable for surfacing high-impact people who are not publicly visible.

    The guests emphasize that many of the most influential operators, builders, and contributors do not have polished public profiles, yet they drive outsized results. This is important because conventional hiring and networking systems tend to over-index on self-promotion. A reputation-first model can reveal talent that would otherwise be missed, improving how teams, founders, and investors discover people.

  • 04

    The Cosign Awards are meant to curate meaningful talent, not establish a competitive leaderboard.

    Rather than treating the awards as a strict ranking system, the organizers use them to identify a broad pool of impactful people across categories like product, engineering, design, and operations. That distinction matters because it frames the product as a discovery and community-building mechanism, not just a status game. This approach may encourage broader participation and reduce the drawbacks of overly gamified rankings.

  • 05

    Cosign is designed to create long-term serendipity that can lead to careers, startups, and partnerships years later.

    The long-term vision is that professional communities formed through Cosign will produce meaningful relationships over time, not just immediate introductions. This is significant because network value often compounds slowly, leading to future cofounder matches, hiring opportunities, and collaborations. The episode frames networking as infrastructure for future ecosystems rather than a transactional contact list.

Cover art for Decoder with Nilay Patel: Can Anthropic afford to put safety first?

Can Anthropic afford to put safety first?

Decoder with Nilay Patel by Vox Media Podcast Network · 37 min

Listen to episode →

This episode focuses on Anthropic’s IPO and what it means for the company’s identity as a safety-first AI lab. Rather than fixating on valuation, the discussion centers on whether Anthropic can keep investing in AI safety while also meeting the growth and profitability expectations of public shareholders. The episode also explores the company’s enormous compute costs, which could force it to grow revenue very aggressively over the next decade. Another major theme is Anthropic’s enterprise strategy: its reputation for trust and data sovereignty is an asset, but customers remain highly price-sensitive and may prefer cheaper or open-source models. Finally, the conversation examines Anthropic’s unusual governance structure and whether it truly protects safety or mostly leaves practical control with the founders.

Key Takeaways

  • 01

    Anthropic’s IPO is really a test of whether a safety-first AI company can survive public-market pressure.

    The central tension is not the stock price itself, but whether Anthropic can continue prioritizing AI safety once it is accountable to shareholders. Going public increases pressure to deliver growth, margins, and returns, which could conflict with slower or more cautious safety investments. This matters because Anthropic has built much of its brand around being a responsible frontier AI developer.

  • 02

    Compute is Anthropic’s biggest operational constraint, and it will likely drive the company’s financial strategy.

    The episode highlights estimated infrastructure costs approaching $900 billion over time, showing how expensive frontier model training and inference can be. That scale of spending means Anthropic must keep expanding revenue just to fund chip purchases, cloud capacity, and ongoing model development. Concepts like model training, inference, and compute capacity are therefore core to understanding the company’s future.

  • 03

    Anthropic’s public debut marks a landmark moment for the AI industry, not just for the company.

    As the first pure AI lab to go public, Anthropic’s IPO signals how central frontier AI has become to capital markets. The strong market reception reinforces that investor enthusiasm for AI remains intense, even amid debates about risk and regulation. This makes the IPO a bellwether for how markets value AI labs that combine technical ambition with a safety narrative.

  • 04

    Anthropic’s enterprise appeal depends on trust, but customers still make decisions based on cost and flexibility.

    The company’s sales pitch leans heavily on being a 'safe, trustworthy AI' provider with stronger data sovereignty and enterprise-grade controls. However, the episode notes that many businesses are trying to route work to cheaper proprietary or open-source alternatives when possible. That means Anthropic’s brand may help win deals, but pricing pressure could limit how much it can monetize its reputation.

  • 05

    Anthropic’s governance structure is unusual, but its practical effect on day-to-day control may be limited.

    The episode discusses the trust-based governance model and the idea that it is designed to protect the company’s mission. Still, the transcript suggests founders retain most of the real control, while the trust mainly influences board composition rather than operational decisions. This raises an important question about whether corporate governance can genuinely safeguard AI safety or mainly serves as a symbolic check.

  • 06

    There is uncertainty about whether customers will always demand frontier models or settle for 'good enough' AI.

    Anthropic appears to assume that many customers will pay for the best possible model, but the episode challenges that assumption. In practice, enterprises often optimize for cost, speed, and acceptable quality rather than the most advanced frontier model. That debate is crucial because it affects revenue growth, product strategy, and the long-term economics of large language models.

Questions answered in this issue

Why does Kush think compute needs its own market structure?

He argues that compute behaves like a tradable commodity: many suppliers sell similar GPU capacity and many buyers demand it with different constraints. That creates the need for standardized pricing, indices, futures, and hedging tools.

What is the main risk in the neocloud business model?

The core risk is tenant credit quality: neoclouds may lend compute to startups or AI companies that cannot ultimately pay for it. If demand weakens, they could be left holding depreciating hardware and exposed financing.

Why does Cosign think professional profiles are insufficient?

The guests say profiles only reveal so much, especially on platforms like LinkedIn. They believe the more important signal is the quality of the work someone has done and who can vouch for it.

How are people using AI agents with Cosign?

Users are connecting agents through MCP and telling them to search for people, opportunities, or relevant updates daily. That makes Cosign useful as a continuously running discovery layer rather than a one-time search tool.

Why does Anthropic’s IPO matter beyond the stock price?

It matters because the IPO forces Anthropic to balance safety, growth, and shareholder expectations at the same time. The transcript argues that the real issue is not the valuation itself, but whether the company can keep prioritizing safety once public-market pressure intensifies.

How much growth does Anthropic need to satisfy public-market expectations?

One estimate cited in the conversation says Anthropic may need roughly a 17-fold revenue increase over 10 years. That illustrates the scale of commercial expansion required to finance compute and deliver returns comparable to other tech companies.

Explore

Related podcasts