How AMP PBC is Turning AI Compute Into a Public Utility

The artificial intelligence industry in 2026 is choking on its own success. Despite massive leaps in algorithmic capabilities and model efficiency, the physical infrastructure required to build these systems has become prohibitively expensive and scarce. Hardware lead times for enterprise-grade chips have stretched to anywhere from three to seven months, and the cost of renting high-end hardware has surged dramatically. For early-stage founders, the primary barrier to entry is no longer just recruiting top-tier machine learning talent; it is the sheer, brute-force capital required to secure computing power. Without guaranteed access to servers, even the most innovative startups are finding themselves paralyzed.[5]

Enter Anjney Midha and his new venture, AMP PBC. Backed by $1.3 billion in initial funding commitments from heavyweights like Andreessen Horowitz and Y Combinator, AMP is attempting to fundamentally rewire how artificial intelligence companies access their most critical resource. Midha’s proposition is as ambitious as it is necessary: he wants to turn graphics processing units (GPUs) into a public utility, functioning much like a modern electric grid. Instead of hoarding hardware, companies would simply plug in and draw the power they need.[1][2]

“The world runs on electricity, but nobody builds their own power plant,” Midha’s firm argues. Yet, in the current AI ecosystem, startups are effectively being forced to build their own power plants just to keep the lights on. To secure competitive hourly rates from major hyperscale cloud providers, founders often have to sign rigid, multi-year contracts. This dynamic locks up tens of millions of dollars in venture funding before a company even has a clear product-market fit or a stable user base.[2][6]

To solve this, AMP PBC operates as a public benefit corporation with a dual mandate: managing a decentralized compute grid and running a venture operation. The grid is the core technological product, designed to pool underutilized GPUs from independent data centers, research labs, and enterprise operators into a single, shared network. Member AI labs can tap into this massive pool on demand, paying dynamic rates that AMP actively manages to keep below the highly volatile spot market prices charged by traditional cloud vendors.[2]

This utility model directly addresses a massive structural inefficiency in AI development: the stark difference between training and inference workloads. Training a frontier AI model requires massive, predictable bursts of compute running continuously over several months. It is a capital-intensive process, but the parameters are known. Inference—the ongoing process of running the trained model to answer user queries or generate content—is much cheaper per individual request but highly unpredictable, fluctuating wildly based on daily consumer demand and product adoption.[6]

When startups buy or rent GPUs in bulk to train a new model, they are often left holding massive excess capacity once the initial training run is complete. If their product does not immediately generate overwhelming consumer demand to soak up that inference capacity, those expensive chips sit idle, burning through cash. AMP’s grid solves this by allowing operators with excess capacity to route it back into the network, ensuring that hardware is constantly utilized globally and startups only pay for the exact compute cycles they consume.[2][6][7]

The urgency behind this utility model stems from a brutal hardware market that has only tightened in 2026. GPU pricing is rising across all major product lines, driven not just by demand, but by severe memory bottlenecks. Specifically, global shortages in high-bandwidth memory (HBM) and GDDR components have constrained manufacturing. Over the past two years, these memory modules have shifted from being supporting components to becoming the primary cost drivers in GPU production, forcing manufacturers to raise prices across the board.[5][7]

Consequently, cloud rental prices have spiked to punishing levels. Hourly rates for next-generation hardware, such as the Blackwell B200, have reached as high as $4.89 to $7.50, roughly doubling since the start of the year. For ambitious startups attempting to build foundational models or complex autonomous agents, GPU compute now routinely consumes a staggering 40% to 60% of their entire technical budget in their first two years of operation. This financial strain is forcing many promising teams to abandon their research or sell early.[2][7]

Midha witnessed this crisis firsthand during his previous tenure as a general partner at Andreessen Horowitz (a16z). There, he spearheaded the “Oxygen” program, a $1.25 billion initiative designed specifically to guarantee GPU access for the firm’s portfolio companies. He realized that in the modern AI race, capital alone was entirely insufficient; startups were routinely being deprioritized by hyperscale cloud providers who favored massive, multi-billion-dollar enterprise clients. Without guaranteed compute, a startup's bank balance was effectively meaningless.[6]

This realization led to AMP’s second, more disruptive arm: treating compute as a literal financing layer. AMP is not just renting out servers; it is investing compute capacity directly into startups in exchange for equity. In February 2026, AMP participated in Anthropic’s massive $30 billion valuation round, writing a $300 million check that underscored its position as a new kind of Silicon Valley power broker. They are proving that server access is now a fundamental part of the capital stack.[3]

“A smaller check with guaranteed capacity can beat a larger check that leaves a lab stuck in a cloud waiting room,” notes recent industry analysis of AMP’s strategy. By bundling server access into the venture capital model, AMP is arguing that the next great AI breakthroughs will be limited not by human ingenuity, but by who can get enough compute at the exact right moment. This approach gives their portfolio companies a distinct operational advantage over rivals who must navigate the open market.[3]

However, this model introduces profound questions about power, centralization, and gatekeeping in the AI industry. While AMP pitches its grid as a democratizing force that lowers the barrier to entry, skeptics worry it could simply replace one set of corporate gatekeepers with another. If compute becomes the primary financing layer, the venture firms controlling the grid will wield extraordinary influence over which companies are allowed to ship products and which are starved of the resources needed to survive.[3][7]

There is also a growing geopolitical dimension to the compute crisis. As artificial intelligence becomes deeply integrated into national defense, healthcare systems, and economic infrastructure, governments are increasingly viewing compute capacity as a critical sovereign asset. The divide between “hypercenters”—regions with abundant hardware and cheap energy—and “compute deserts” is widening rapidly. AMP’s decentralized grid could theoretically help bridge this gap by routing capacity globally, but it remains tethered to the physical realities of data center locations, export controls, and local energy grids.[7]

Despite these long-term uncertainties, the shift toward a utility model appears inevitable for the industry's continued survival and growth. The current trajectory—where every ambitious startup must hoard chips, over-provision their infrastructure, and negotiate aggressively with hyperscalers—is financially unsustainable for the broader ecosystem. It creates a chilling effect on innovation, where only the most heavily funded teams can afford to experiment. By abstracting the hardware layer and pooling global resources, AMP hopes to eliminate this friction and return founders’ focus to where it truly belongs: algorithmic innovation, product design, and solving real-world problems.[1][7]

If Midha’s $1.3 billion experiment succeeds, it will fundamentally alter the venture capital landscape and the underlying mechanics of software development. Compute will no longer be viewed as a static, burdensome line item on a monthly cloud bill; it will become the very currency of the AI economy, flowing dynamically and efficiently to wherever the most promising ideas are being built. In doing so, this utility model may finally fulfill the original promise of the tech industry: making frontier artificial intelligence accessible to anyone with a groundbreaking idea and the drive to execute it.[2][3]