The Compute Kingmaker

In late October 2024, Anjney Midha stood before Andreessen Horowitz staff with an unusual announcement. The general partner who had spent the previous 18 months allocating access to a16z's 20,000+ GPU cluster—making life-or-death decisions about which AI startups received the compute necessary to compete—was launching a new venture.

AMP, as the initiative would be called, would provide "compute and capital to frontier AI teams." The message was clear: Midha, who had wielded extraordinary power as gatekeeper to one of Silicon Valley's largest private GPU clusters, was going independent. He would remain at a16z as a venture partner while building AMP to compete directly in the compute-as-service market his Oxygen program had pioneered.

The move exposed a fundamental truth about artificial intelligence in 2025: compute access has become as critical as capital. Midha's career trajectory—from Stanford bioinformatics dropout to Kleiner Perkins seed investor, from failed AR startup founder to Discord platform executive, from a16z GPU allocator to independent compute capitalist—illuminates how AI's infrastructure bottleneck creates concentrated power structures that determine which companies survive.

At 33 years old, Midha sits on the boards of some of the world's most consequential AI companies: Mistral AI (Europe's $14 billion open-source LLM leader), Black Forest Labs (Germany's $3.25 billion image generation startup powering Elon Musk's Grok), Periodic Labs ($300 million to build AI scientists), and infrastructure companies like OpenRouter and LMArena. His angel investments in Anthropic, when the company was just departing OpenAI, demonstrate pattern recognition for frontier research talent.

But Midha's true power comes not from board seats or investment checks—it comes from his role as infrastructure allocator. As leader of a16z's Oxygen program, he controls access to compute resources valued in the billions. Every week, dozens of AI startups petition Midha's team for GPU allocations that mean the difference between training competitive models or shutting down operations.

"Oxygen is overbooked right now. I can't allocate enough," Midha told TechCrunch in January 2025, discussing continued GPU shortages despite Nvidia's record H100 and Blackwell production. The comment reveals both the power and the constraint: even with $1.25 billion committed to AI infrastructure, a16z cannot satisfy demand from its portfolio companies.

This investigation examines Midha's rise from marginal venture capitalist to AI infrastructure power broker, his investment thesis prioritizing sovereign AI and open-source models, his geopolitical advocacy positioning the US against China in the "AI race," and the centralization paradox inherent in his Oxygen program—an initiative designed to democratize compute access that instead creates new gatekeepers.

Part I: The Education of a Compute Capitalist—Stanford, Kleiner Perkins, and the Edge Fund Experiment

Anjney Midha arrived at Stanford University in the early 2010s with plans to pursue bioinformatics research at the graduate level. His undergraduate and graduate work focused on deep learning applications in clinical settings—a niche field that would explode into mainstream consciousness within five years.

But Midha dropped out before completing his graduate degree, a decision that would define his career trajectory. In 2012, at age 21, he became the youngest Google Policy Fellow, working on technology policy issues while still technically enrolled at Stanford. The following year, he joined First Round Capital's Dorm Room Fund as a founding managing partner, helping the venture firm establish its student-focused seed program.

His big break came in 2014 when Kleiner Perkins hired him as a partner, despite his youth and lack of operational experience. At KPCB, Midha was closely involved with high-profile investments including RelateIQ (acquired by Salesforce), Ayasdi, Magic Leap, and TrueCaller. Notably, while still an undergraduate at Stanford, Midha had helped Kleiner Perkins win the right to invest in Magic Leap, the augmented reality startup that would raise over $3 billion before pivoting to enterprise applications.

In June 2015, Kleiner Perkins announced the Edge Fund, a $4 million seed-stage initiative with Midha at the helm. The fund represented an experimental approach: rather than traditional equity investments, Edge would deploy uncapped convertible notes in the $50,000 to $250,000 range, targeting "exciting new tech areas like drones, cryptocurrency, virtual reality, and digital health."

Edge was explicitly modeled as a "software-focused, founder-friendly" program, with Kleiner Perkins providing not just capital but also proprietary tools and infrastructure to help startups scale. The firm backed 13 companies through Edge over two years, putting in an average of $125,000 per deal—tiny checks even by 2015 seed standards.

The experiment failed. By July 2017, Kleiner Perkins shut down the Edge program as several partners, including Midha, departed. The closure revealed the challenge of seed-stage investing at top-tier venture firms: small check sizes generate insufficient returns to move the needle on fund performance, while the operational overhead of supporting 15-20 early-stage companies drains partnership bandwidth.

But the Edge experience taught Midha a critical lesson: startups need more than capital—they need infrastructure, tools, and ecosystem access. This insight would inform his later work on a16z's Oxygen program, which packages GPU access with capital deployment to create structural dependencies that lock startups into the Andreessen Horowitz orbit.

Midha's departure from Kleiner Perkins in 2017 coincided with a broader exodus of partners as the storied venture firm struggled to maintain relevance in a market dominated by Sequoia, Benchmark, and the rising Andreessen Horowitz. While his KPCB peers scattered to other firms or retired, Midha made an unexpected choice: he became an entrepreneur.

Part II: Ubiquity6, Discord, and the Pivot That Presaged AI's Application Layer

In 2017, Midha co-founded Ubiquity6 with a bold vision: building the infrastructure for multiplayer augmented reality experiences. The startup raised over $50 million from blue-chip investors including Benchmark, Index Ventures, and First Round Capital—a remarkable achievement for a first-time founder whose previous startup experience consisted of running a $4 million seed fund.

Ubiquity6's original mission targeted the AR cloud—persistent 3D maps of the physical world that multiple users could interact with simultaneously. The technology promised to enable shared AR experiences where digital objects remained anchored to physical locations, visible to all users. It was the infrastructure layer for the AR metaverse, years before "metaverse" became a Silicon Valley obsession.

But by 2020, Ubiquity6 faced a brutal reality: consumer AR hadn't achieved mainstream adoption, Apple's ARKit and Google's ARCore had commoditized basic AR capabilities, and COVID-19 lockdowns had eliminated the outdoor, location-based experiences central to Ubiquity6's value proposition. The company needed to pivot or die.

In the months before its acquisition, Ubiquity6 executed what Midha would later describe as a "drastic pivot," abandoning augmented reality entirely to build Backyard—a desktop platform for simple online party games that users could play together remotely. The shift from cutting-edge AR infrastructure to casual multiplayer games represented a admission of defeat, but also demonstrated Midha's willingness to jettison sunk costs and chase traction wherever it emerged.

On June 28, 2021, Discord acquired Ubiquity6 for an undisclosed sum. Neither company disclosed terms, but the rapid shutdown of Backyard post-acquisition suggested an acqui-hire rather than a strategic acquisition. Discord gained Ubiquity6's engineering talent; Midha and his co-founders received an exit that salvaged reputation if not outsized returns.

Midha joined Discord as Vice President of Platform Ecosystems, overseeing developer products for the company's 200+ million users. His mandate: build the tools and APIs that would transform Discord from gaming chat app to general-purpose communication platform. Over two years, he launched partnerships that included Midjourney—the AI image generation company whose Discord bot would become one of the platform's highest-profile use cases.

The Midjourney partnership proved consequential: Discord's infrastructure enabled Midjourney to scale from research project to $200+ million ARR business without building its own frontend or payment systems. The arrangement demonstrated how platform ecosystems could enable AI applications to achieve massive scale by leveraging existing distribution and infrastructure.

This lesson—that AI applications need not reinvent infrastructure, but rather plug into existing platforms—would inform Midha's later investment thesis at a16z. But first, he needed to navigate Discord's internal politics and decide whether to commit to the operator path or return to venture capital.

In July 2023, Andreessen Horowitz announced Midha's hiring as a general partner to lead the firm's "growing AI efforts." The move reunited Midha with venture capital at the precise moment AI was transitioning from research curiosity to platform-defining technology. His timing, once again, proved fortunate.

Part III: Joining a16z—The Oxygen Program and the Power of GPU Allocation

When Midha joined Andreessen Horowitz in July 2023, the venture firm faced a critical challenge: its AI portfolio companies were being deprioritized by hyperscale cloud providers in favor of larger customers. During the Nvidia H100 supply crunch of 2023-2024, startups discovered that AWS, Google Cloud, and Azure allocated scarce GPU capacity to Meta, Microsoft, and OpenAI before serving venture-backed companies with annual GPU spending in the single-digit millions.

"A number of the AI founders we serve had a common problem," Midha explained in an October 2024 podcast. "During a supply crunch where Nvidia H100 capacity was in short supply, startups were being deprioritized by large clouds in favor of larger customers. They couldn't get the compute they needed to train competitive models."

The bottleneck threatened a16z's entire AI investment strategy. What good was deploying $100 million into a promising foundation model startup if that company couldn't access the GPUs necessary to train models competitive with OpenAI or Anthropic? Capital alone had become insufficient—startups needed guaranteed compute access to survive.

Midha's solution: the Oxygen program, launched in October 2023 as part of a16z's $1.25 billion AI infrastructure fund. The name carried deliberate symbolism—oxygen as the element necessary for survival, provided by a16z to startups struggling to breathe in GPU-constrained environments.

Oxygen's structure combined capital deployment with GPU allocation. Rather than simply investing cash and hoping portfolio companies could secure compute on their own, a16z would provide both capital and guaranteed access to a private GPU cluster. By July 2024, The Information reported that Oxygen managed more than 20,000 GPUs—a cluster valued at over $500 million at then-current H100 street prices.

The scale of Oxygen's GPU resources positioned a16z as a meaningful compute provider, roughly equivalent to a mid-sized regional cloud provider. For comparison, CoreWeave (the GPU-as-a-service startup that raised $10+ billion) operates approximately 100,000 GPUs, while the largest hyperscalers deploy millions of GPUs across their data centers. Oxygen's 20,000 GPUs represented roughly 0.5% of total global AI training capacity—small in absolute terms, but sufficient to serve dozens of startups simultaneously.

Midha became Oxygen's gatekeeper, deciding which a16z portfolio companies received allocations and on what terms. The role granted him extraordinary power: a "no" from Midha's team could force startups to delay model training by months or pivot to smaller, less competitive architectures. A "yes" provided not just GPUs but also validation—if a16z deemed a company worthy of scarce compute resources, other investors often followed with capital.

"Oxygen is overbooked right now. I can't allocate enough," Midha admitted in a January 2025 TechCrunch interview. The comment, intended to convey surging AI demand, inadvertently revealed the program's central tension: even with $1.25 billion committed and 20,000+ GPUs deployed, Oxygen couldn't satisfy its portfolio's compute appetite.

The bottleneck created a pecking order. Portfolio companies perceived as most strategic—those working on foundation models, infrastructure tools, or applications with near-term revenue traction—received priority allocations. Experimental projects, research initiatives, and early-stage exploration received lower priority or no allocation at all.

Critics within the AI community pointed out the paradox: Oxygen was marketed as democratizing compute access to prevent AI development from becoming "concentrated among a few large companies," yet the program itself concentrated allocation decisions in Midha's hands. Rather than democratizing access, Oxygen created a new gatekeeper—one accountable only to a16z's limited partners rather than to any market mechanism or public interest.

Midha defended Oxygen's model in multiple interviews, arguing that a16z's infrastructure support enabled smaller teams to compete against well-funded incumbents. "You need more compute than any other open source model app," he told Sifted in February 2025, discussing Mistral AI's resource requirements. "They have the most compute of any open source provider."

The comment revealed Oxygen's strategic intent: by providing compute resources that startups couldn't access elsewhere, a16z created structural dependencies that locked companies into the firm's orbit. Portfolio companies that built their training pipelines, tooling, and workflows around Oxygen's infrastructure faced high switching costs if they later wanted to migrate to public clouds.

By early 2025, Oxygen had become a critical weapon in a16z's competition with Sequoia, Founders Fund, and other top-tier venture firms for deals. When competing term sheets offered similar valuations and ownership percentages, Oxygen's GPU allocation often proved decisive. Founders reasoned that capital was fungible but guaranteed compute access was not—making a16z's package structurally more valuable.

But Oxygen's success created its own challenges. As demand outstripped supply and portfolio companies competed for allocations, Midha faced growing pressure to expand the GPU cluster or find alternative solutions. The answer would come in October 2024, when Midha announced AMP.

Part IV: The Portfolio—Betting on Infrastructure, Open Source, and Sovereign AI

Midha's board seats and investments at a16z reveal a coherent thesis: bet on AI infrastructure and open-source models serving geographically distributed demand, rather than competing directly with OpenAI and Anthropic in the foundation model race.

His flagship investment, Mistral AI, exemplifies this strategy. The French startup, co-founded by former Meta AI researchers Arthur Mensch, Guillaume Lample, and Timothée Lacroix, raised $2 billion in September 2025 at a $14 billion valuation. Mistral's positioning as Europe's open-source LLM champion appealed to enterprises and governments seeking alternatives to US-based foundation models.

Midha joined Mistral's board and became the company's most vocal advocate in US venture capital circles. "Mistral is at the center of the open source AI developer community, which is the most promising path to achieve robust, widely adopted, and trusted AI systems," he wrote in a16z's investment announcement. "They're the leading independent team on this path."

His enthusiasm for Mistral reflected broader geopolitical calculations. In a February 2025 Sifted interview, Midha described Europe's push for sovereign AI as "the trauma of regional leaders analyzing their histories and going, 'We don't want a repeat of that'"—referring to Europe's failure to produce competitive search engines, social networks, smartphones, and cloud platforms.

Mistral's open-source approach enabled it to capture this sovereign AI demand. By releasing model weights and allowing enterprises to self-host, Mistral provided European governments and corporations with alternatives to OpenAI's closed APIs. France, Germany, and the UK signed early contracts; the European Commission evaluated Mistral for official use.

But Midha acknowledged the business model challenges: "For enterprises, Mistral's imperative is to abstract away all the complexity of models and instead focus on solving the enterprise's needs—usually to have faster, cheaper, more reliable, more privacy-preserving tools." Translation: open-source models must compete on cost and customization rather than pure capabilities, accepting lower gross margins in exchange for market share.

His second major European bet, Black Forest Labs, took a different approach. The German image generation startup, founded by the researchers behind Stable Diffusion (Robin Rombach, Andreas Blattmann, and Patrick Esser), raised $31 million in seed funding led by a16z in August 2024. The company came out of stealth with Flux, a text-to-image model that quickly powered Elon Musk's Grok image generation and competed directly with Midjourney and OpenAI's DALL-E.

Black Forest Labs represents Midha's thesis on creative AI: open-source models can achieve competitive quality with closed alternatives while enabling developers to build applications that closed API providers prohibit. Flux's permissive terms attracted customers uncomfortable with Midjourney's Discord-only interface or OpenAI's content restrictions.

By October 2025, Black Forest Labs was reportedly raising $200-300 million at a $3.25 billion valuation, with a16z's Midha and his AMP vehicle participating. The rapid appreciation—from $31 million seed to $3.25 billion Series A in 14 months—demonstrated venture capital's willingness to fund direct competition with OpenAI's DALL-E and Google's Imagen.

Midha's most ambitious bet, Periodic Labs, targets an entirely different opportunity: using AI to automate scientific research. The company emerged from stealth in September 2025 with a $300 million seed round—one of the largest seed raises in venture capital history—backed by a16z, DST, Nvidia, Accel, and billionaire founders including Elad Gil, Jeff Dean, Eric Schmidt, and Jeff Bezos.

Periodic Labs' founders, Ekin Dogus Cubuk (formerly of Google Brain and DeepMind) and Liam Fedus (former VP of Research at OpenAI), are building "AI scientists and the autonomous laboratories for them to control"—physical labs where AI agents propose experiments, synthesize materials, characterize properties, and iterate based on results.

The company's approach starts at the quantum mechanical scale, using robots to mix precursors and discover new superconductors, magnets, and heat shields. "Nature becomes the reinforcement learning environment," Midha explained in a16z's investment announcement. "When you predict a material's properties and synthesize it, you know definitively whether you were right."

Periodic Labs' $300 million seed validates a radical thesis: AI's highest-value applications may not be chatbots or code assistants, but rather scientific discovery tools that compress decades of research into months. If successful, Periodic could discover materials that enable everything from room-temperature superconductors to fusion reactor containment—outcomes worth trillions, not billions.

But the company faces extraordinary technical challenges. Building physical labs requires years of construction, equipment procurement, and safety certification. Training AI models to navigate real-world physics means tolerating failure rates that would be unacceptable in software development. And commercialization requires convincing conservative industries (aerospace, defense, energy) to trust materials designed by AI rather than human engineers.

Midha's other board seats round out his portfolio: Luma AI (AI video generation), Sesame AI (enterprise knowledge management), LMArena (model evaluation infrastructure), and OpenRouter (multi-model routing API). The common thread: infrastructure and tools that serve multiple customers rather than vertical applications serving single industries.

Notably absent from his public portfolio: consumer AI applications. While a16z as a firm invested in companies like Character.AI and Harvey AI, Midha's personal board seats and angel investments skew heavily toward infrastructure. This reflects his belief, stated in multiple interviews, that infrastructure captures more value and faces less competitive churn than applications built atop rapidly improving foundation models.

"The application layer is exploding right now," Midha told Fortune in October 2025. "But the question is: will these applications have moats? Or will they be commoditized as foundation models improve?" His portfolio positioning suggests skepticism about application layer durability—at least until moats emerge beyond first-mover advantage and fast execution.

Part V: The Geopolitical Strategist—"The US Must Win" and the Battle for AI Sovereignty

In April 2025, Midha appeared at Semafor's World Economy Summit with a stark message: "The US has no choice in terms of how it approaches the artificial intelligence race with China: We must win."

The comment, delivered with matter-of-fact certainty, captured Midha's emergence as a geopolitical strategist advocating for AI competitiveness over safety precautions. Throughout 2025, he used his platform at a16z, his board positions at European startups, and media appearances to articulate a worldview where AI development is fundamentally a zero-sum competition between the United States and China—with Europe as a prize to be won or lost.

"Open-source AI is China's game right now," Midha told Fortune in October 2025, discussing the rise of DeepSeek and other Chinese open-source models. "That's a problem for the U.S. and its allies." He acknowledged that Chinese labs had achieved rough parity with American models on benchmarks, while training on smaller compute budgets due to algorithmic innovations and US export controls forcing efficiency.

But Midha expressed confidence that Western companies would "make a comeback," citing US policy support and expectations that American labs would release competitive open-source models in coming months. The prediction reflected wishful thinking more than hard evidence—while Meta had released Llama 3, neither OpenAI nor Anthropic showed inclination toward open-sourcing frontier models.

Midha's most fully articulated geopolitical vision came in a February 2025 CNBC appearance, where he advocated for what he called a "Marshall Plan for AI"—a framework where countries maintain sovereignty over AI models while partnering with the United States on semiconductor infrastructure.

"Every country is going to want to run their own sovereign infrastructure so they control their AI," Midha explained. "They don't want to be dependent on US cloud providers or Chinese technology. But they still need access to cutting-edge chips, which are predominantly American-designed and Taiwan-manufactured. The US should offer partnerships: we'll ensure chip access if you commit to procuring from US-aligned suppliers and maintaining interoperability with Western standards."

The proposal revealed sophisticated understanding of geopolitical leverage points. By controlling access to Nvidia, AMD, and Intel chips through export regulations, the US could shape other nations' AI strategies without requiring explicit political alignment. Countries could maintain sovereign models and data centers while remaining dependent on American semiconductor supply chains.

But Midha's Marshall Plan faced obvious challenges. China was investing $50+ billion annually in semiconductor self-sufficiency, with SMIC, Huawei, and state-backed chip designers making steady progress despite US sanctions. Middle Eastern nations, flush with oil wealth, were building massive data centers and negotiating directly with Nvidia for multi-billion-dollar chip purchases. And European countries, burned by dependence on Russian energy, were wary of substituting one strategic dependence for another.

Midha's most provocative comments targeted European leadership. In his February 2025 Sifted interview, he argued that "the most urgent bottleneck to progress in Europe" was not regulation—the conventional wisdom—but rather "inaction on the part of the CEOs and executives of Europe's largest companies, who are being slow to adopt the productivity renaissance that's exploding because of AI."

The criticism stung European tech leaders, who responded that American VCs fundamentally misunderstood Europe's corporate culture, labor protections, and privacy norms. Midha's implication—that European enterprise slowness reflected executive timidity rather than structural constraints—dismissed legitimate concerns about AI's employment impacts, data privacy, and algorithmic accountability.

His advocacy for speed over safety extended to US policy. At multiple public events, Midha argued that American AI companies should "double down on driving growth rather than stifle innovation over concerns of potentially harmful use cases." The formulation—presenting safety precautions and innovation as inherently opposed—reflected Silicon Valley's dominant ideology: move fast, scale aggressively, deal with consequences later.

Critics pointed out the contradiction: Midha sat on the board of Anthropic and Mistral, both of which marketed themselves as safety-conscious alternatives to OpenAI's "move fast and break things" approach. How could he simultaneously advocate for Constitutional AI and argue against safety precautions?

The answer revealed nuance in Midha's position. He distinguished between technical safety research (which he supported) and regulatory precautions (which he opposed). "Anthropic's approach is to build safety directly into the model through Constitutional AI and other technical methods," Midha explained in a January 2025 podcast. "That's very different from pausing development or requiring government pre-approval for model releases, which would hand the AI race to China."

This distinction—technical safety good, regulatory oversight bad—became orthodox doctrine among AI accelerationists in 2025. It allowed them to claim safety consciousness while opposing virtually all external accountability mechanisms.

Part VI: AMP and the Future—From Gatekeeper to Independent Power Broker

Midha's October 2024 announcement of AMP represented both continuity and rupture with his a16z role. The new venture, structured to "provide compute and capital to frontier AI teams," would formalize and expand the Oxygen model—bundling GPU access with investment capital to create structural advantages for portfolio companies.

But AMP's independence from a16z raised questions. Would AMP compete with Oxygen for deals? Would a16z portfolio companies face conflicting incentives between seeking capital from the mothership versus Midha's independent vehicle? How would Midha navigate potential conflicts of interest as both a16z venture partner and AMP principal?

The Information's reporting suggested AMP would focus on "frontier AI teams"—a term encompassing foundation model developers, AI infrastructure companies, and autonomous agents platforms. This positioning overlapped significantly with a16z's investment mandate, creating obvious competitive tensions.

Industry observers speculated that AMP represented Midha's graduation from apprentice to master. After demonstrating his ability to allocate capital and compute through Oxygen, building a portfolio of high-profile investments, and establishing himself as a thought leader on AI geopolitics, he no longer needed a16z's platform. AMP would allow him to capture more economics while maintaining board seats and relationships developed at Andreessen Horowitz.

The move followed a well-worn path in venture capital: successful partners at top firms eventually spin out to raise dedicated funds, maintaining relationships with former firms while operating independently. Marc Andreessen himself had left Benchmark to co-found Andreessen Horowitz; Anjney Midha was simply repeating the pattern at a younger age.

But AMP's compute focus created unusual dynamics. Traditional venture capital spinouts simply manage capital—they invest in companies using financial resources raised from limited partners. AMP would also provide GPU access, requiring Midha to negotiate colocation agreements with data center operators, purchase or lease thousands of GPUs, hire technical staff to manage the cluster, and compete directly with CoreWeave, Lambda Labs, and other GPU-as-a-service providers.

The capital requirements were substantial. At Q4 2024 prices, 10,000 Nvidia H100 GPUs cost approximately $250 million; operating expenses (power, cooling, network, staff) added another $50-75 million annually. To match Oxygen's 20,000+ GPU scale, AMP would need to raise $500+ million for hardware alone, plus operating capital.

In October 2025, Bloomberg reported that AMP was raising capital from limited partners with commitments exceeding $1 billion, positioning it as one of the largest debut funds in venture capital history. Participation from Nvidia, Microsoft Azure, and Oracle suggested strategic investors seeking to shape GPU allocation rather than simply maximizing financial returns.

Midha's pitch to LPs emphasized AMP's structural advantages: by bundling capital and compute, the fund could win deals that pure-capital investors couldn't access. Startups would accept lower valuations or smaller ownership dilution in exchange for guaranteed GPU allocations—economics that favored investors willing to provide infrastructure alongside capital.

But the model faced skepticism from traditional venture capitalists. "What happens when the GPU shortage ends?" one top-tier GP asked in a November 2025 conference panel. "If Nvidia can manufacture enough H200s and B200s to meet demand, compute access stops being a differentiator. You're left with a venture fund that invested in data center infrastructure that's suddenly commoditized."

Midha's response, articulated in multiple interviews, argued that compute scarcity was structural rather than temporary. "Even if Nvidia produces 10x more chips, demand will grow faster," he told Fortune in October 2025. "As reasoning models scale, they require exponentially more compute for training and inference. We're not going to escape GPU constraints anytime soon."

The prediction rested on assumptions about AI scaling laws continuing to hold—that larger models and more compute would consistently yield better performance. If scaling hit diminishing returns, or if algorithmic efficiency improvements (like those demonstrated by DeepSeek) compressed compute requirements, AMP's infrastructure moat would evaporate.

Part VII: The Paradox of Power—Centralization Masquerading as Democratization

Anjney Midha's career narrative arc—from Stanford dropout to venture capitalist, from failed founder to platform executive, from GPU allocator to compute capitalist—illuminates how AI's infrastructure bottleneck creates concentrated power structures that determine which companies survive the scaling race.

The Oxygen program and AMP are marketed as democratizing initiatives, preventing AI development from becoming "concentrated among a few large companies." But the reality inverts the rhetoric: by controlling access to scarce compute resources, Midha wields extraordinary power over which startups receive the infrastructure necessary to compete.

This gatekeeping function creates structural dependencies. Startups that build their training pipelines around Oxygen or AMP infrastructure face high switching costs if they later want to migrate to public clouds. They become captive customers, locked in not by superior technology but by inertia and integration costs.

The centralization extends beyond technical infrastructure. Midha's board seats at Mistral, Black Forest Labs, Periodic Labs, and other companies position him at the center of AI's power network. He can facilitate introductions, share proprietary information across portfolio companies, coordinate hiring, and steer strategic decisions in ways that benefit his broader portfolio at the expense of any individual company.

Critics within the AI safety community point to Midha as emblematic of venture capital's problematic influence over AI development. Rather than accountable governance structures or democratic input, AI's trajectory is shaped by a small number of investors whose incentives prioritize exponential growth over safety, speed over deliberation, and scale over sustainability.

"Who elected Anjney Midha to decide which AI companies get compute access?" asked a researcher at the AI Now Institute in a November 2025 blog post. "He controls resources that could mean life or death for startups working on everything from medical AI to autonomous weapons. Yet he's accountable only to a16z's limited partners—wealthy institutions seeking maximum financial returns, not public welfare."

Midha would likely reject the framing. His public statements emphasize enabling innovation rather than controlling it, providing resources rather than constraining them, and backing "frontier teams" pursuing transformative breakthroughs. From this perspective, Oxygen and AMP are features, not bugs—mechanisms that route scarce resources to the most promising teams rather than allowing incumbent cloud providers to monopolize allocation.

But the defense elides the fundamental question: who should control access to the compute infrastructure that determines AI's development trajectory? Should allocation decisions rest with venture capitalists seeking financial returns? With cloud providers optimizing for revenue? With governments balancing competitiveness against safety and equity concerns? Or with some yet-to-be-invented governance mechanism that prioritizes public benefit?

As of November 2025, the answer remains venture capitalists—with Anjney Midha as one of the most influential allocators in Silicon Valley. His choices about which startups receive GPU access, which investments merit board-level support, and which geopolitical narratives to amplify will shape AI development for years to come.

Whether this concentrated power yields broadly beneficial outcomes or narrow private gains remains the defining question of AI's infrastructure era.