Human and Systemic Bottlenecks Drive AI Superintelligence Progress

In a world racing towards AI superintelligence, Anj Midha, a prominent AI investor and founding investor in Anthropic, argues that the true bottlenecks are not algorithmic, but deeply human and systemic. This conversation reveals the hidden consequences of misaligned incentives and the critical need for infrastructure standardization, particularly compute. Midha’s insights are crucial for founders, investors, and policymakers who need to navigate the complex landscape of AI development, offering a strategic advantage by highlighting where to focus investment and effort for sustainable, long-term progress rather than chasing fleeting trends. Anyone building or investing in AI will gain clarity on the foundational elements required for genuine advancement.

The Unseen Bottlenecks: Beyond Algorithmic Hype

The prevailing narrative around AI often focuses on the next breakthrough model or the latest algorithmic innovation. However, Anj Midha, a seasoned investor with deep ties to Anthropic and a founder himself, argues that this focus is dangerously myopic. The real impediments to reaching superintelligence, and the true sources of competitive advantage, lie not in the code itself, but in the underlying systems and human factors that enable its development and deployment. Midha identifies four critical bottlenecks: context feedback, compute, capital, and culture. While algorithmic innovation was once a significant hurdle, he posits that it has largely become a function of culture. A strong, mission-driven culture attracts top researchers who, unburdened by architectural dogma, naturally drive algorithmic progress.

The more pressing challenges, in Midha’s view, are context feedback and compute. Context feedback refers to the unique, real-world data generated from deploying models in specific domains. This is where significant commercial advantage lies, as demonstrated by his own venture, Periotic Labs, which uses robots and physical verification to generate novel material science data, a domain where existing LLMs falter due to a lack of specialized data.

"Wherever there are some unique context feedback loops that are missing today, that's where you probably have the biggest bottlenecks on capabilities. And so what you should be doing if you're trying to advance the frontiers is going, 'Okay, you know, these models suck.'"

This highlights a critical downstream effect: the hype around "AI for science" often outpaces reality because the necessary specialized data is locked away in proprietary systems or academic labs. Midha’s solution for Periotic Labs--building a physical lab to generate this data--illustrates a profound consequence: true innovation in specialized domains requires vertical integration and a willingness to invest heavily in creating proprietary feedback loops, a path many are unwilling to tread due to its immediate cost and complexity. This creates a durable moat for those who do.

The Compute Crisis: A Pre-Standardization Era

The bottleneck of compute is perhaps the most tangible and systemically critical. Midha draws a powerful analogy to the Industrial Revolution, comparing the current state of AI infrastructure to pre-standardization electricity in 1885. Compute, unlike electricity, is not yet fungible. Different chip architectures and even different generations of the same chip family cannot seamlessly work together, leading to vast amounts of "stranded compute"--billions of dollars of GPUs sitting idle because they cannot be integrated into broader, efficient systems.

"Compute is not fungible today. So forget fungibility of compute across different manufacturers, like Nvidia and AMD. Within a manufacturer, Nvidia chips, for example, H100s, the GB200s, the GB300s, these are all completely different chip types. So if you have one cluster where you're doing a training run on H100s, and then you want to continue post-training of that, do a distributed training run of that training workload on GB200s, doesn't work."

This lack of standardization creates an "infrastructure wastage crisis." The consequence of this non-fungibility is that companies that could otherwise scale their AI capabilities are starved of the necessary computational resources. This directly impacts innovation, as frontier models are developed faster than the chips that can efficiently run them. Midha’s initiative, the Amp Grid, aims to act as an "independent system operator" for compute, much like the electricity grid, coordinating capacity to ensure teams can access what they need without over-provisioning. This is a long-term investment, requiring significant capital and a systems-level approach to design, but it promises to unlock massive downstream value by enabling broader access to compute. The immediate discomfort of building this infrastructure is precisely why it creates a lasting advantage; few are willing to undertake such a complex, capital-intensive endeavor.

Sovereign Data, Sovereign AI: The Geopolitical Dimension

The discussion around context feedback loops and compute leads directly to geopolitical considerations. Midha highlights the concept of "sovereign data" and its implications for AI development, particularly in Europe. The US Cloud Act, which allows US authorities access to data held by US companies, creates a significant barrier for mission-critical workloads in sensitive sectors within Europe. This necessitates local AI infrastructure and models.

The emergence of companies like Mistral AI, which are building European-based AI infrastructure and models, is a direct response to this need. Midha's investment thesis in Mistral is rooted in "independence at scale at every part of the AI infrastructure stack." This move towards sovereign AI stacks is not just about data privacy; it's about maintaining technological independence and ensuring that critical AI capabilities are not solely concentrated in a few global hubs. The consequence of this trend is a potential fragmentation of the AI landscape, but also an opportunity for regional innovation and a more distributed ecosystem. The early days of Anthropic, where 21 out of 22 VCs said "no" to their ambitious capital-raising plans, underscore the difficulty of investing in frontier technology before its value is widely understood. Midha’s conviction, rooted in his understanding of the underlying technical and systemic challenges, allowed him to see the potential where others saw only risk.

"I invested a bunch of my money that was just life savings, which was not much, given I was a poor founder at the time where most of my net worth was tied up in Discord stock. And it pains me sometimes to look back at the emails of friends. So I introduced them to 22, you know, friends up and down Sand Hill Road. So there's some investors there, and we got 21 no's. And I was like, 'What, what are you guys thinking?'"

This experience highlights a crucial aspect of competitive advantage: the ability to identify and invest in foundational, often unglamorous, infrastructure and capabilities before they become obvious market trends.

The "Frontier Systems" Shift and the Future of Venture

Midha challenges the conventional categorization of companies like Anthropic as merely "foundation model companies." He argues they are, and always have been, "frontier systems companies." This distinction is critical because it reframes the entire approach to building and investing in AI. A frontier systems company focuses on the entire loop of research, development, deployment, and feedback, not just the core model. This requires a different kind of capital allocation--one comfortable with massive, long-term capital expenditures ("CapEx") in businesses that aim to dominate entire categories.

This shift has profound implications for venture capital. Midha advocates for a "back-to-the-future" era of venture capital, characterized by deep, hands-on partnerships between investors and founders, akin to the early days of Intel or Genentech. This contrasts sharply with the more passive check-writing prevalent in recent years. The value accrual, he suggests, will increasingly come from co-founding and incubating businesses, rather than simply funding them.

"The commercial community has forgotten how to build businesses, and they've forgotten the difference between first principles and marketing. That's the problem. That's one of the other misalignment problems. The ground truth of these businesses, machine learning systems businesses, they've always been frontier systems businesses. They were never just foundation model businesses."

The consequence of this shift is that VCs who fail to adapt--those who are not "building with AI" themselves or deeply understanding the technical nuances--will struggle. The future of venture lies in identifying and backing those who can navigate the systemic bottlenecks and build enduring "frontier systems."

Key Action Items:

• Immediate Actions (Next 1-3 Months):

  • Educate Yourself on Bottlenecks: Deeply study the four core bottlenecks (context feedback, compute, capital, culture) as identified by Anj Midha. Understand how they apply to your specific domain or investment thesis.
  • Assess Compute Fungibility: For teams building with AI, analyze your current compute infrastructure. Identify areas where compute is non-fungible or underutilized due to lack of standardization.
  • Seek Proprietary Data Loops: If operating in a specialized domain, actively explore or build unique context feedback loops that competitors cannot easily replicate.
  • Review Capital Allocation: For investors, scrutinize your portfolio for companies that are merely chasing algorithmic trends versus those building foundational infrastructure or addressing systemic bottlenecks.

• Medium-Term Investments (Next 3-12 Months):

  • Invest in Standardization Efforts: Support or advocate for open standards in compute and AI infrastructure. This could involve contributing to open-source projects or backing companies focused on interoperability.
  • Explore Vertical Integration: For companies, consider strategic vertical integration where it creates defensible moats, particularly in data acquisition and model deployment.
  • Develop Mission-Driven Culture: Foster a culture that prioritizes long-term mission alignment over short-term gains, attracting and retaining top talent capable of driving innovation.
  • Prioritize Deep Partnership Models: For VCs, shift towards more hands-on, co-founding, or incubation models, especially for frontier technology companies.

• Longer-Term Investments (12-24 Months+):

  • Build or Access Compute Infrastructure: For significant AI undertakings, proactively secure access to scalable and ideally fungible compute resources. Consider participating in initiatives like the Amp Grid.
  • Champion Sovereign AI Development: If relevant to your region or market, support the development of local AI ecosystems and infrastructure to ensure independence and resilience.
  • Focus on "Frontier Systems" not just "Models": Reframe your investment strategy and company building efforts around the entire "frontier systems" approach, encompassing the full lifecycle of AI development and deployment.
  • Embrace "Optimal Competition": As investors, seek to back a small number of leading teams in critical areas rather than spreading capital thinly across many undifferentiated competitors. This requires patience and conviction.