AI compute demand is so intense that even older-generation GPUs are becoming more expensive to rent, signaling a fundamental shift in how software will be priced and packaged. According to Anish Acharya, a partner at venture capital firm Andreessen Horowitz (a16z), the AI market is experiencing "infinite demand and severely constrained supply," which is reshaping everything from how enterprises choose AI models to what consumers will pay for software.

What's Driving the Shift in AI Software Pricing?

The clearest sign of AI's explosive demand is counterintuitive: B200 GPUs, which are not even the latest generation of graphics processing units, are seeing their hourly rental prices rise instead of fall. In a normal technology cycle, computing hardware becomes cheaper over time. This price increase directly demonstrates how scarce AI compute has become and how desperately companies need it.

This supply crunch is forcing a rethinking of software economics. Acharya noted that consumer willingness to pay for software has undergone a fundamental transformation. "Unlike the 99-cent app era of the past, today's consumers are willing to pay $200 a month," he explained. This shift is giving rise to what he calls "luxury software," a new category of premium AI-powered applications that command prices previously unthinkable for consumer software.

Acharya

"If the ceiling for software used to be $20, what should a $200 or even a $2,000-per-month SKU of your product look like?" Acharya asked, encouraging founders to rethink product positioning accordingly.

Anish Acharya, Partner at Andreessen Horowitz

How Are Large Language Models Becoming Less Interchangeable?

Acharya pushed back against the widespread belief that large language models (LLMs), the AI systems that power tools like ChatGPT, are becoming commodities. Instead, he argued that these models are diverging significantly in their capabilities and even their "personality traits." OpenAI's models excel at knowledge work, while Anthropic's Claude leans toward software engineering. Some models behave in rigid, neurotic ways suited for accounting tasks, while others are more open and creative, better suited for design work.

This divergence means enterprises need a more strategic framework for selecting which AI model to use. Acharya proposed an approach based on the "value ceiling" of each business role, which determines whether a company should invest in expensive, cutting-edge models or fine-tuned open-source alternatives.

Steps for Enterprises to Choose the Right AI Model

• Uncapped Upside Roles: For positions like sales and product management where output can scale indefinitely, use top-tier closed-source models regardless of cost. Acharya explained that "it is economically rational to pay any price for a model that is even 1 IQ point smarter" because you cannot predict how much revenue a new feature or major client might generate.
• Capped Upside Roles: For back-office functions like finance and human resources where accuracy is the goal, use fine-tuned open-source models to optimize cost efficiency. As Acharya noted, "you can't close the books 10 times better than accurate," so premium models offer diminishing returns.
• Integration Considerations: Evaluate how easily each model integrates with existing systems. Integration moats, which once protected enterprise software giants, are now vulnerable to disruption by coding agents that can automate migration and system changes.

Which Corporate Advantages Are Actually Safe From AI Disruption?

Acharya offered a nuanced view of which business advantages will survive the AI revolution. Traditional moats like network effects, scale effects, and brand effects remain robust. "No amount of coding agents will make Nike stop being Nike," he noted. "Instagram's power was never about the complexity of building the app, but about the network behind it".

Acharya

However, one type of advantage faces existential threat: the "integration moat." Enterprise software companies like SAP have historically locked in customers by making it extremely difficult and expensive to switch to competitors or migrate to new versions. But coding agents that can automate software migration are undermining this protection. "Even migrating from one version to the next now carries existential risk," Acharya warned.

Despite this threat, Acharya maintained a measured perspective on enterprise software disruption. Enterprise software spending currently accounts for only 8 to 12 percent of total spending, and core systems like payroll or customer relationship management (CRM) systems demand such high precision that pure AI coding agents are unlikely to fully disrupt them in the near term. "The upside isn't particularly large, but the downside risk of getting it wrong is essentially unlimited," he explained.

Acharya

Why Are Startups Suddenly Finding Opportunity in the Application Layer?

Contrary to earlier concerns that large AI model providers would dominate the entire market, Acharya observed that major model providers are actually moving in the opposite direction. Instead of moving up the stack to control applications, they are integrating downward into inference compute, the computational work required to run trained models. This leaves significant opportunity for application-layer startups.

The reason is straightforward: inference workloads are highly standardized, making it easy to build scale advantages. The application layer, by contrast, involves extremely complex pricing, packaging, and meeting diverse customer needs. "Turning 'intelligence primitives' into economic outcomes for specific industries like credit unions requires a completely different product form," Acharya explained. This operational complexity is precisely where startups can compete effectively.

In the consumer market, Acharya believes AI applications are experiencing something akin to "the Christmas when the iPhone launched in 2009." The core barrier that previously prevented consumer AI breakout was high customer acquisition cost. When he attempted to develop an AI-powered X app, acquiring a single new user cost as much as $250. Today, low-cost, high-performance open-source models have changed that dynamic dramatically.

Acharya

This shift is also changing the profile of successful founders. Acharya observed that startup teams now feature "fewer and fewer MBAs, and more and more researchers." These young founders with purely technical backgrounds may lack business maturity, but they possess extremely high technical acuity and carry no intellectual baggage. "They believe anything is possible," he noted. On the capital side, this also explains why nine-figure mega seed rounds are now emerging. With the leverage of AI tools, a small number of top-tier talents paired with sufficient capital can cover a product portfolio that previously required a massive team to complete.

Acharya