Inference Is the New AI Stack

Cheap Chinese models are driving investor anxiety because they appear to threaten the pricing power of frontier U.S. labs like OpenAI and Anthropic.

The fear is that lower-cost alternatives could compress margins and weaken the business case for premium AI models. However, the episode suggests that model price alone does not determine market winners; access to inference capacity, distribution channels, and enterprise adoption still matter heavily.

AI is no longer just a sector-level story; it has become a major macroeconomic force.

The episode cites AI investment as contributing a large share of U.S. GDP growth and a significant portion of S&P 500 returns and earnings growth. That makes AI spending relevant not just to startups and labs, but to investors, index performance, and the broader economy.

Concerns about circular financing, capex blowouts, and spending outrunning revenue are real, but they have not yet proved a bubble exists.

The segment acknowledges the classic warning signs of speculative excess: heavy capital expenditures, financing loops, and pressure for revenue to catch up. Still, the host argues that the evidence so far looks more like rapid infrastructure buildout than irrational mania, especially because demand remains strong.

The real constraint in AI is not just cheaper model training; it is the ability to serve demand at scale through inference capacity and distribution.

Even if model creation becomes cheaper, companies still need enough compute, infrastructure, and go-to-market reach to monetize usage. This distinction helps explain why low-cost models do not automatically collapse the economics of premium AI providers.

Token caps and usage limits suggest a more disciplined AI market, but they do not necessarily indicate stalled growth.

The episode frames token caps as signs that companies are managing costs and throttling usage, rather than evidence that demand has disappeared. In practice, the limits appear small relative to the size of the opportunity, leaving room for further expansion.

Periodic market freakouts may actually be healthy because they act like pressure valves that slow excess before it becomes a true speculative bubble.

Instead of allowing euphoria to compound unchecked, repeated fears about pricing, spending, and competition force investors and companies to reassess assumptions. The episode’s core thesis is that this reflexive skepticism may be one of the reasons the AI market has not yet become a runaway bubble.

Inference speed, not just model quality, is becoming the defining bottleneck in AI product performance.

Feldman emphasizes that users experience AI through latency, so even a powerful model can feel unusable if it responds too slowly. The key concept is "tokens per second per user," which connects raw compute to real product experience. This matters because faster inference enables more interactive, more useful AI applications and changes what kinds of products can be built.

AI hardware scaling is constrained by a few critical supply-chain chokepoints, especially HBM, advanced packaging, and 3nm wafer capacity at TSMC.

The episode highlights that the AI boom is not only a software story but also a manufacturing and logistics story. These bottlenecks limit how quickly the industry can scale compute, regardless of how much demand exists. Understanding these constraints is essential for evaluating who can actually deliver AI capacity at scale.

Cerebras' wafer-scale architecture is designed around minimizing memory movement, which Feldman sees as the main limiter for inference workloads.

Instead of relying on conventional GPU designs, Cerebras uses a giant chip with substantial on-chip SRAM to reduce the need to shuttle model weights back and forth. This architectural bet is significant because inference often depends more on memory bandwidth and latency than on raw arithmetic throughput. The discussion shows how chip design can be optimized for specific AI workloads rather than generalized compute.

As AI systems become agentic and reasoning-heavy, CPU demand rises because the system must plan, verify, and take actions around the model itself.

Feldman frames the accelerator as the "brain" and the CPU as the "body," reflecting a shift from simple text generation to end-to-end task execution. This matters because agents do not just answer once; they chain multiple steps, checks, and external tool calls. That increases the value of low-latency infrastructure and makes inference efficiency even more important.

Cerebras' competitive moat comes from architectural uniqueness rather than easily copied features or brute-force spending.

Feldman argues that competitors cannot simply outspend their way into the same design advantages because the architecture itself changes the performance equation. This is important in a market where many players can buy similar GPUs but few can reproduce a wafer-scale system. The takeaway suggests that deep hardware differentiation can create durable defensibility in AI infrastructure.

AI model progress is moving so quickly that today's best models may soon look outdated, and younger engineers are already adapting workflows around AI-native tools.

Feldman suggests that current capabilities will soon seem primitive, which implies a very short product and platform half-life in AI. He also notes that younger engineers use AI differently, hinting at a shift in how teams build, debug, and ship software. This is significant because it suggests the competitive advantage may increasingly come from adapting workflows, not just adopting tools.

Ask DoorDash is driving real behavior change by making delivery discovery more natural through conversational, natural-language queries.

The shift from traditional search to an agentic commerce interface lowers friction for restaurant and grocery discovery, especially when users have vague or nuanced needs. This matters because better discovery can increase conversion and broaden the range of purchases people make on the platform. It highlights the value of AI interfaces that understand intent rather than forcing rigid keyword-based input.

The AI interface is not just a novelty; it is already influencing purchase behavior, including more first-time restaurant orders and larger grocery baskets.

According to the founders, roughly half of the trajectories involve customers ordering from places they had never ordered from before, and grocery basket sizes are reportedly about 40% larger. That suggests the interface is creating measurable incremental demand rather than simply shifting existing orders around. For DoorDash, this is important because it links AI product design directly to revenue and marketplace expansion.

DoorDash approaches autonomy as a use-case-first problem, which led it to build Dot and its delivery autonomy stack in-house.

Rather than starting with a generalized robotics platform and then looking for applications, DoorDash began with the delivery problem and designed technology around it. This approach is significant because delivery has specific economic and operational constraints that require purpose-built solutions. It also reflects a product-management mindset: solve the customer and logistics problem first, then engineer the system to fit.

Delivery autonomy is likely to be multimodal, with different vehicle types serving different delivery contexts.

The founders argue that no single autonomous vehicle will handle every terrain, distance, or environment efficiently. Instead, the future may include a mix of robots, human Dashers, and potentially other delivery modes depending on the situation. This is significant because it moves the conversation away from winner-take-all autonomy toward practical system design and operational flexibility.

DoorDash’s real-world delivery and drop-off data gives it an advantage in solving the 'first and last 100 feet problem.'

The hardest part of autonomous delivery is often not the road segment, but the exact handoff at the merchant and customer endpoints, where real-world variation is highest. DoorDash has unique data on neighborhoods, drop-offs, merchant interactions, and edge cases that can improve autonomy systems in ways generic mapping data cannot. This makes its logistics footprint a strategic asset for robotics and AI development.

The founders believe autonomy will expand the market and may actually lead to more Dashers, not fewer.

Their thesis is that cheaper, more efficient delivery will increase total demand, creating more overall delivery volume. In that scenario, autonomous systems complement human labor rather than fully replacing it. This is a notable counterpoint to common fears about automation and frames autonomy as market expansion rather than labor substitution.