When Tokens Become Capital: Anand Chandrasekaran on the Work AI Must Actually Deliver

1. Guest Background

Anand Chandrasekaran joins The Futurists with experience on several sides of company building: as an operator, a co-founder and a venture investor at Celesta. That combination gives this interview its subject and perspective. The discussion examines how spending on AI infrastructure might become useful business activity, and why financing technology is connected to understanding the work a company must deliver.

Chandrasekaran says he grew up in India, moved to the United States for graduate study at Stanford and accumulated more than twenty years of operating experience before entering venture investing. He describes a path through General Catalyst and then Celesta. An earlier platform transition helps explain his interest in expanding access. Recalling Yahoo around 2013–2014, he contrasts US traffic growth supported by paid partnerships and affiliates with products in China and India sometimes adding five million to ten million users a week without spending on traffic acquisition. He interprets that difference as consumers moving directly into mobile internet use.

His contact-center experience supplies a second reference point. He says Five9’s revenue run rate rose from approximately $150 million to $600 million during the pandemic, and that Zoom subsequently made an acquisition offer. These are his recollections, rather than independently established results or evidence of AI-driven growth. Their relevance is that he had operated within a software-delivered service model before considering what generative AI might change about the underlying tasks.

2. What the Episode Covers

The episode begins with a question about the destination of capital: after infrastructure is built, what generates the value that could justify it? The host cites an unnamed report forecasting $11 trillion in AI infrastructure investment over multiple years, while acknowledging that some spending may never occur. He then mentions $7 trillion diverting capital from other industries without explaining how the figures relate. These numbers establish the conversation’s scale, but remain attributed framing. The episode does not provide a basis for combining them or treating either as a settled investment total.

Chandrasekaran responds with Celesta’s distinction between investing below and above large language models. Below the model sit hardware, memory, interconnects and other components of data-center capacity. Above it sit applications in contact centers, construction, radiology and supply chains. The division matters because the obstacles differ. Infrastructure investment addresses the availability of computational resources; application investment must establish how a particular industry can use those resources to deliver work. He therefore emphasizes combining AI specialists with people who understand the relevant domain.

Crescendo illustrates the change he wants to pursue. He recalls sitting at General Catalyst in January 2023 and recognizing ChatGPT as a breakthrough that could move software beyond making employees more productive toward taking over some tasks in their entirety. That judgment became a rationale for building the company. According to his account, co-founder Andy Lee had spent twenty-five years running a business with approximately $2 billion in revenue before becoming Crescendo’s CEO. General Catalyst and Celesta supported the venture. The significance of Lee’s experience is operational: Chandrasekaran wanted someone familiar with delivering an entire service involved in deciding how AI should perform it.

High Octave, also called HiO in the conversation, extends the approach to small-business growth. Chandrasekaran describes an ambition to provide enterprise-grade AI capabilities to smaller businesses. He says the company is led by co-founder Golan, formerly Wix’s COO, who spent seven years building that business. The interview uses this second venture to explore whether owners can acquire execution capacity without becoming experts in another layer of business software themselves.

The broader structure is Chandrasekaran’s three-stage account of AI development: remove infrastructure bottlenecks, transform industries through cooperation among AI, domain and transformation specialists, and enable faster discovery. He also says demand exceeds supply and that balancing the two will take time. This framework links the episode’s subjects without proving that one stage automatically produces the next. Building capacity, delivering reliable tasks and achieving wider social benefits remain separate accomplishments.

3. Core Views: Reasoning, Examples, and Limits

The useful idea behind “tokens are the new capital” is that AI capacity has competing uses. Chandrasekaran describes a cycle in which compute becomes tokens, tokens perform labor and tasks, and profits from automation purchase more compute. To illustrate allocation, he recounts a discussion attributed to Satya Nadella: should Microsoft direct scarce capacity toward Office products, model development or cloud services sold to enterprises? Those uses draw on a common resource while pursuing different returns. The Nadella discussion is secondhand within this episode. More fundamentally, the reinvestment cycle is a proposed economic relationship, not an investment calculation. No cost structure or return model establishes that producing more tokens necessarily creates more profit.

HiO gives that resource argument a practical object. Chandrasekaran describes owners who know their products but lack the labor or skills to create content, respond to changing customer demographics or turn a first order into repeat purchasing. He proposes an AI growth partner that finds, adjusts and executes those loops while the owner concentrates on customers and products. The important distinction is between obtaining a tool and obtaining completed work. His further suggestion that ambition becomes the principal constraint is a product vision, however. The interview provides no task-completion rates, failure-handling evidence or growth attribution that would establish the claim. Execution capacity can expand without establishing that customers will buy the resulting offer.

Rob’s HubSpot story introduces a separate measure of value: whether the customer can operate independently. He recalls a company he co-founded moving through four or five service tiers. The platform cost less than additional hiring, but the business failed to develop some internal capabilities and he could not see a route to leaving. This is not evidence that HiO has the same characteristics. It explains why access and portability require different questions. Chandrasekaran replies that his service unlocks activities owners previously could not undertake and stresses end-to-end execution. That addresses the reason to purchase, but the exchange contains no specific answer about data migration or an exit mechanism. A service can create useful capacity while leaving its customer increasingly dependent.

The labor discussion exposes another distinction that the growth narrative cannot settle. Rob describes expansion without additional hiring as technological change favoring capital, then asks whether automation threatens the career ladder associated with Indian BPO and IT outsourcing. Chandrasekaran responds that AI can handle notes, summaries and back-office work, allowing contact-center employees to solve complex problems and deepen their product knowledge. That is an account of how an existing worker’s task mix might improve. It does not establish the number of future jobs, the availability of entry-level opportunities or the costs of moving workers into different roles. Individual advancement and aggregate employment are related questions, but one is not evidence for the other.

Outcome-based contact-center pricing offers a more explicit mechanism. Chandrasekaran argues that separating activity from results could make ten-, twenty- or thirty-minute interactions across phone, chat and email more viable. Instead of treating every additional minute primarily as a cost, a company might value the opportunity to understand customers. His manufacturing example imagines a defective batch generating customer calls the following day, accelerating feedback. The reasoning is plausible as a mechanism, but the episode supplies no measured comparison. It also leaves the settlement rules unspecified: what counts as an outcome, who verifies it and how responsibility is assigned. Longer conversations and more contact could support better service; neither demonstrates that better service occurred.

Discovery and education require the same discipline. Chandrasekaran describes an unnamed portfolio company combining 3D-printed human lymph nodes with AI testing loops, reportedly shortening drug-testing time and running a pipeline approaching that of a major pharmaceutical company. The account suggests experimental speed and parallelism, but provides no precise reduction, clinical success evidence or regulatory approvals. It cannot establish successful medicines or lower patient prices. Likewise, his vision of opening university curricula through AI, supplemented by the host’s examples of age-adjusted explanations and quizzes, concerns access and interaction. The episode contains no learning-outcome evaluation. Across these examples, the central unresolved passage is from greater technical supply to reliable delivery and then to benefits experienced by the intended recipient.

4. Learning and Application

For an investor or business operator, the token framework becomes useful when attached to one task with an identifiable recipient and acceptance standard. The following practices are analytical applications of the interview, rather than reported customer results. Trace the resources consumed, the work completed, the evidence of acceptance and the value someone will pay for. Include checking, corrections and exception handling when comparing the service with an existing process. Chandrasekaran’s reinvestment cycle depends on revenue or savings exceeding computing and delivery costs. His acknowledgment of power, energy and chip constraints also means an application forecast should not assume that every increase in demand can immediately obtain additional capacity.

A small business considering a growth partner could begin with a bounded assignment, such as developing follow-up communication intended to encourage repeat orders. Specify what the owner approves, what the service executes and how the result will be assessed. Then test continuity: export customer information and approved materials, and ask whether another person could operate the process. This responds to the dependence raised by Rob’s account and left unresolved in the interview; it does not presume that HiO lacks export or migration capabilities. The tradeoff becomes clearer when convenience and portability are evaluated together, including the owner’s time saved and the practical cost of switching.

For contact-center procurement, define the outcome before changing the billing unit. A pilot could specify when an issue counts as resolved, which cases require a person, how later corrections affect acceptance and how disagreements are handled. Longer conversations may be permitted without becoming the performance target. To test the manufacturing-feedback mechanism, examine whether defect reports are accurately recorded, reach the responsible team and lead to a response. Otherwise a company risks replacing a visible activity cost with a promise whose meaning differs between buyer and supplier. The interview offers a reason to explore outcome pricing, but not a ready-made contract or a demonstrated price advantage.

Chandrasekaran’s operating approach also suggests a bounded way for investors to participate. He says some partners have served as temporary CEOs for three months, six months or a year during company transitions, arguing that close involvement reduces the need to catch up and coordinate through board meetings. A practical adaptation would define the problem, decision authority, handover milestones and end of the assignment. Time spent inside a company should still be connected to a specific contribution. Similarly, his advice that young people join AI-native companies, loosen their attachment to titles and help shape the technology can guide a search for learning opportunities. Candidates should examine access to complete tasks, supervision and explanations of failures; a growth mindset alone does not supply those resources.

Finally, preserve separate evidence standards for ambitious applications. A faster drug-testing loop calls for further evidence about clinical performance and regulatory progress. An AI tutor that explains and quizzes calls for checking whether the learner actually understands and retains the material. Neither endpoint was demonstrated here. Using the episode well means turning its promises into questions that can be answered at the next stage, while keeping technical capability, operating reliability and final benefit distinct.

Source

• Original episode: The Futurists - EPS_353: Tokens Are The New Capital with Anand Chandrasekaran