In-depth Interview with Perplexity CEO: Recursive Flywheel, Edge-Cloud Hybrid Orchestration, and the Surge of Small and Medium Enterprises

AI Deep Dive | The Real Bet of Perplexity: When a Company is Left with Only One Person, Where Will the Profits Flow?

1. The "recursive effect" describes customer expansion but has yet to prove that the platform has a moat.

Customers make money with AI and then use more income to purchase AI services, which is indeed more commercially valuable than relying on novelty to maintain activity. However, it is essential to distinguish between two things: customers becoming increasingly dependent on such tools and customers being unable to leave a particular supplier.

Network effects mean that new users can increase the value of use for other users; the "recursive effect" mentioned in interviews is closer to the growth of tool consumption driven by customer business growth. It can lead to revenue expansion but will not automatically prevent competitors from taking customers away with lower prices.

What needs to be observed is whether the workflows, business memories, permission configurations, and feedback accumulated by customers allow the platform to understand their business better; whether switching platforms requires revalidating the entire production process.

Therefore, judging such companies cannot rely solely on user numbers or computing power consumption. It is necessary to track old customer retention, usage growth after deducting the impact of price reductions, and the gross profit generated from servicing these customers. The deeper the usage, only converting it into sustainable profits and migration costs constitutes a business barrier.

2. "A one-person company earns $3 million a year," primarily reveals changes in cost structure.

The program publisher recorded Srinivas's statement: an independent entrepreneur earns about $3 million a year through an automated newsletter business while spending over $1 million on Perplexity. This case comes from management's narrative and cannot be regarded as an audited business sample, nor can the income be understood as net profit. Sourcery

According to this set of numbers, platform spending alone exceeds one-third of the income, and the remaining portion must cover customer acquisition, data, auditing, refunds, and the labor of the operator. The so-called "one-person company" actually hands over parts of production activities completed by employees to external computing platforms.

This change reduces the fixed costs of hiring and management but increases dependence on supplier prices and service stability. Assuming income remains at $3 million, if platform spending rises from $1 million to $1.5 million, with all other conditions unchanged, the operating surplus would decrease by $500,000.

Therefore, what entrepreneurs need to manage is not just "human efficiency" but also the computational cost of each delivery, customer renewal rates, and profit tolerance after supplier price increases.

3. The organizational advantages of small companies may accompany further concentration of the entire industry infrastructure.

AI allows small teams to reduce internal communication levels, but computing, models, payments, identity verification, and customer entry may still be concentrated in the hands of a few large platforms. The increase in the number of enterprises does not equate to the decentralization of economic power.

One can understand an AI-native small company as: the front end is controlled by entrepreneurs who manage customers, while the back end purchases production capacity from platforms. Whether it can retain profits depends on whether customers are purchasing interchangeable automated outputs or unique industry judgments, channels, and trust.

Amazon already positioned AWS as a new business aimed at software developers in its 2006 shareholder letter, while emphasizing that new investments must meet capital return requirements. Its business insight is that helping others lower the entrepreneurial threshold can itself become a massive infrastructure business.

Perplexity hopes to occupy a similar position, but it must also procure from underlying model and computing suppliers. For entrepreneurs, the key question is: have they accumulated customer assets, or have they merely secured a high-income job temporarily on a particular platform?

4. AI increases productivity, but that does not mean businesses using AI can retain all the profits.

If an organization uses AI to reduce the cost of research reports from 1,000 yuan to 200 yuan, it may initially expand profits. However, when competitors gain access to the same tools, prices may continue to decline, and cost savings ultimately flow to customers.

Therefore, production efficiency, corporate profits, and social welfare must be observed separately. Reports become cheaper, services become more widespread, and society still benefits; however, this benefit may not necessarily be reflected in higher valuations for suppliers.

A study released by NBER in 2023, utilizing data from 5,179 customer service personnel, found that generative AI assistance increased the number of problems solved per hour by an average of about 14%, with new and low-skilled employees benefiting more significantly. This proves that specific jobs can be improved but does not directly imply that corporate profits will simultaneously grow by 14%.

For small and medium-sized enterprises, it is worthwhile to enter businesses that "still have differentiation after efficiency improvements": exclusive data, clear responsibilities, stable channels, and complex customer relationships. Purely relying on the advantage of production speed is often the easiest to be consumed by competition.

5. "Models are all contributors" is a strategically narrative with vested interests.

From the perspective of the orchestration platform, having more upstream models with increasingly similar capabilities is, of course, a good thing. Competition among suppliers can lower procurement costs and reduce the platform's dependence on any one model.

However, upstream manufacturers also want to capture customers and can directly provide search, document processing, agents, and enterprise connectors. They are both suppliers and potential channel competitors. The term "contributor" does not change the reality of both parties competing for profits.

The value of neutral orchestration lies in taking on the complexity of selection, testing, switching, and failure handling for customers. If it merely forwards requests to different models, the price difference space can easily be compressed; if it can stably complete cross-system tasks, customers then have reason to pay for the entire delivery capability.

Therefore, it should be examined whether the platform maintains quality and cost stability after switching underlying models. Neutrality must be realized as the ability to replace suppliers, verify results, and continuously reduce costs to be worth charging.

6. Whether model price reductions can improve profits depends on the total resources consumed to complete tasks.

A cheaper response does not equate to a cheaper workflow. Agents may search more pages, call more tools, initiate more subtasks, and conduct longer reviews due to enhanced capabilities.

Assuming the cost of each model call decreases by half, but the number of calls to complete a task increases to three times the original, the total cost of the model may actually rise by 50%. Adding in browser, storage, third-party data, and retries, cost changes may contradict the advertised unit price trends.

A more useful business metric is: how much it costs to complete a task that meets acceptance criteria. This metric includes success rates, review labor, and redo costs.

Charging based on usage can protect platform gross margins but may cause customer budgets to spiral out of control; fixed subscriptions are beneficial for sales but may attract high consumption, low-profit users. A long-term viable charging method must align platform revenue growth with the actual value obtained by customers, rather than rewarding ineffective calculations.

7. Local computing is a capital allocation choice and cannot be compared solely based on cloud bills.

Perplexity's official documentation confirms that its products can run local models on supported Mac, AMD, and NVIDIA hardware. However, device support does not equate to establishing a strategic partnership with device manufacturers, and one cannot assume that Apple's cooperation has been realized solely based on Mac compatibility. Perplexity Help Center

Local solutions convert some ongoing service expenditures into device investments, operations, and capacity management. Assuming an additional device investment of $6,000, used for three years, with annual maintenance and power expenditures of $1,000, the average annual cost would be about $3,000, not accounting for capital costs.

If it can only replace $1,000 of cloud expenses annually, it is not cost-effective; if it can stably replace $10,000, and the task quality is comparable, it may be attractive.

For existing idle devices, short-term decisions can focus on incremental costs; for new devices to be purchased, complete costs must be calculated. The cloud and local cost multiples mentioned in interviews can only be meaningfully compared when model capabilities, context, utilization rates, and speed requirements are consistent.

8. Distributed devices can handle some demand but cannot simply add up power to form a supercomputing center.

One hundred million devices each running at 50 watts indeed corresponds to 5 gigawatts of power arithmetically. However, this does not equate to adding 5 gigawatts of power supply, nor does it mean forming equivalent computing power that can be uniformly scheduled.

End devices may sleep, lose connectivity, or be occupied by users, and hardware performance and memory capacity vary; cross-device communication also incurs costs. Their advantages are better suited for independent tasks close to local data and cannot simply be judged by total power to determine whether they can undertake large centralized computing.

The constraints of centralized power supply do exist. A 2026 report from Lawrence Berkeley National Laboratory shows that in areas with data, the median time from application to grid connection for power generation projects completed in 2025 exceeds five years. This statistic pertains to power generation projects and cannot be directly equated to data center electricity waiting times.

A more reasonable judgment is the division of labor between edge and cloud: local processing is suitable for decentralized tasks, while the cloud handles high-intensity inference and large-scale services. Beneficiaries depend on the bottleneck position rather than the label of "decentralization."

9. The security capabilities of agents determine how high-value business they can enter.

AI that only provides suggestions typically has errors that remain at the information level; AI that can modify files, send emails, and operate business systems turns errors into real losses. The greater the authorization, the higher the potential value, and the greater the responsibility.

Related products also need to accurately distinguish: the official codebase describes Bumblebee as a read-only supply chain scanner for checking software packages, extensions, and development tool metadata; Numbat focuses on visibility of terminal agent activities, local detection, optional pre-operation blocking, and post-event reconstruction, and cannot simply be categorized as one managing terminals and another managing servers.

The commercial value lies in turning risk control into a mechanism that customers can check: which files are readable, which actions are executable, who approves payments, and whether failures can be reversed.

Local deployment does not automatically eliminate risks; excessive permissions, malicious inputs, and erroneous operations still exist. What enterprises are willing to pay for is demonstrable control capabilities, not a mere promise of "data not leaving the premises."

10. The cheaper the code, the more valuable the problem definition and acceptance criteria.

Software delivery includes requirement understanding, system design, permission management, testing, and maintenance; writing code is just one part of it. Accelerating one link cannot proportionally shorten the entire process.

Assuming a project spends 40% of its time on coding, with the remaining time on requirements, testing, and coordination, even if coding time is reduced to zero, the total time can at most decrease by 40%. This is a hypothetical calculation but is sufficient to illustrate why project delays may still occur after a surge in code output.

AI will also introduce a new review burden: the more generated functions, the more content maintainers need to understand and verify. Without clear boundaries, automation may amplify rework.

The more scarce capability is defining what is worth doing, what results count as qualified, and how to handle failures. Industry experts may thus gain new entrepreneurial entry points: transforming work that previously relied on experiential judgment into clear processes and verifiable deliveries, rather than first generating a fully functional, vaguely defined software.

11. When enterprises procure AI, they should first calculate cash benefits before discussing how many hours have been saved.

Saving ten hours a week for employees does not mean the enterprise has saved ten hours of wages. Only if this time reduction leads to less overtime, avoids new hiring, increases billable deliveries, or improves customer retention can it translate into financial results.

Assuming a system costs 60,000 yuan annually, with an additional 40,000 yuan for deployment, auditing, and maintenance, the enterprise needs at least 100,000 yuan of verifiable annual benefits to cover costs. If the benefits only come from "feeling more efficient," the return on investment has not yet been established.

Pilot projects should simultaneously record the costs of existing processes, the complete costs after AI implementation, error rates, and final outputs, while retaining comparisons. For growth-oriented businesses, it is also necessary to prevent attributing all sales growth to AI that would have grown anyway.

The first applications should be chosen for high frequency, clear acceptance, and correctable errors. Projects truly worth scaling are those with qualified quality and continuously decreasing unit delivery costs; unattended duration and generation quantity cannot replace this standard.

12. Valuation ultimately depends on remaining cash flow; grand narratives cannot replace capital returns.

The phrase "valuation exceeding $30 billion" in the candidate title should not be adopted directly: the program's publisher used an estimated valuation of about $20 billion in their introduction on October 12, 2026, and noted an annualized revenue of about $750 million. These are merely disclosure figures; annualized revenue does not equal the actual annual revenue achieved, much less profit. Sourcery

Even with rapid revenue growth, the question remains: how much cash can be left after deducting model procurement, computation, support, and R&D? Is revenue growth reliant on subsidies? Will customers continue to use the service after receiving discounts?

Assuming two companies both have revenues of $1 billion, one with a gross margin of 80% and the other 40%, the resources available to cover R&D, sales, and management differ by $400 million. Even with the same revenue multiple, the economic quality can vary significantly.

Whether Perplexity can become a high-value platform ultimately depends on its ability to continuously retain a portion of the revenue that customers are willing to pay and that upstream cannot easily take away. The high efficiency of small businesses must also be validated through customer renewals and profits.

Core judgment: AI is lowering the threshold for organizational production, but it will not automatically lower the thresholds for acquiring customers, building trust, and assuming responsibility. When intelligence can be purchased on demand, long-term profits will tend to favor those who can define demand, validate results, and manage customer relationships.