Anj Midha

Founder of AI compute infrastructure company AMP and science incubation firm Periodic Labs.

Mentions

Appeared in

Discussed in

Key positions and views

Compute is the primary bottleneck for AI progress and the main determinant of success for AI companies, superseding algorithms or capital [6, 11].

Data sovereignty requirements, driven by the US CLOUD Act, are fundamentally fracturing the global cloud market and ending the unchallenged dominance of US hyperscalers [3, 5, 28].

China is competing asymmetrically in AI by focusing on full-stack systems co-design to maximize hardware performance and using adversarial distillation to copy Western model capabilities, rather than competing directly on chip manufacturing [4, 9].

The next major breakthroughs in AI will be in scientific domains like material science, which are seeing 'super-exponential' returns on compute, unlike saturated areas like coding [1, 13, 14].

The traditional venture capital model is ill-suited for funding frontier AI, as evidenced by its widespread failure to invest in Anthropic's seed round and a broader misallocation of capital [7, 18, 20].

Podcast consensus on Midha

Points of consensus

▶Midha consistently argues that access to compute is the primary bottleneck in AI, asserting it will determine the success of inference companies [6] and poses the greatest risk to the revenue targets of major labs like OpenAI and Anthropic [11].Apr 2026

▶He repeatedly identifies data sovereignty, driven by regulations like the US CLOUD Act [3], as the key factor fracturing the cloud market and creating the first major challenge to hyperscaler dominance in 15 years [5, 28].Apr 2026

▶His analysis of China's AI strategy is coherent across multiple claims, focusing on asymmetric tactics like full-stack systems co-design to maximize hardware performance [9, 27] and adversarial distillation to copy Western model capabilities [4].Apr 2026

▶He views AI's application in scientific domains like material science as a key area for future growth, noting that it is currently yielding 'super-exponential gains' [14] but is hampered by a lack of specialized training data [1].Apr 2026

Points of debate

▶Midha's assertion that traditional venture capital firms are failing to capture value in frontier AI, exemplified by their mass rejection of Anthropic's seed round [7, 18, 20], challenges the prevailing narrative of Silicon Valley's investment prowess.Apr 2026

▶His claim of a significant 'GPU wastage bubble' where billions in compute capacity sit idle [17] presents a contrarian view to the more common narrative of extreme and universal GPU scarcity.Apr 2026

▶The proposal for a coordinated 'iron dome' for frontier model inference [12] suggests a level of threat-sharing and cooperation that may be unrealistic in a fiercely competitive commercial landscape.Apr 2026

▶His prediction that major AI labs could fail primarily due to a lack of compute [11] is a bearish take that contrasts with the generally bullish market sentiment surrounding these leading companies.Apr 2026

Key themes

▶The Geopolitics of ComputeApr 2026

Midha frames the AI race as a geopolitical struggle for technological sovereignty, centered on compute infrastructure. He highlights China's strategy of systems co-design [9] and adversarial distillation [4] to catch up, and Europe's need to build a massive, independent compute stack to counter US dominance enforced by policies like the CLOUD Act [3, 10, 28].

For investors and analysts, this theme suggests that AI infrastructure should be viewed not just as a tech sector, but as a strategic national security asset, where data sovereignty will drive significant regional investment and market fragmentation.

▶Compute as the Ultimate ConstraintApr 2026

Midha argues that access to compute is the single most critical bottleneck in the AI industry, superseding capital or algorithms. This scarcity will determine the winners among inference companies [6], poses the greatest threat to the revenue models of major labs like OpenAI and Anthropic [11], and is currently plagued by a 'GPU wastage bubble' [17].

Companies that can secure and efficiently manage massive compute supply, like his own company AMP [21], will hold immense power and value, potentially becoming the new 'kingmakers' of the AI era.

▶AI's Next Frontier: Scientific DiscoveryApr 2026

Midha posits that while AI's gains in text-based domains are hitting diminishing returns [13], its application in science is just beginning to show 'super-exponential gains' [14]. He points to the lack of specialized physics and chemistry data as a key hurdle [1] and showcases his own venture, Periodic Labs, which uses LLMs to discover new materials like superconductors [8].

The most significant long-term value from AI may not be in chatbots or coding assistants, but in solving fundamental scientific problems, creating a new class of AI-native R&D companies.

▶The Failure of Traditional Venture Capital in Frontier AIApr 2026

Midha critiques the traditional venture capital model for its inability to grasp the scale and nature of frontier AI investment. He uses Anthropic's seed round, where 21 of 22 VCs passed [20] and no traditional firms participated [18], as a prime example of this failure. He claims there is a significant misallocation of public capital into VC managers who are missing the most important opportunities [7].

The capital structure for frontier AI may require new models beyond traditional VC, involving public benefit corporations, massive debt financing [30], and strategic partnerships, as demonstrated by his company AMP [22].

Source episodes

Sentiment over time

Not enough data for timeline

Changes over time

Circa 2 years ago

Anthropic attempts to raise a $500M seed round, but after 21 rejections from 22 VCs, re-anchors to a $100M round with no participation from traditional VC firms. This event shapes Midha's view on the inadequacy of traditional venture capital for frontier AI [18, 20, 25].

Circa 1 year ago

Midha notes that major LLMs like Claude and Gemini were performing poorly on scientific analysis benchmarks [19], highlighting the data gap [1] that would inform his focus on AI for science.

Present

Midha identifies a 'GPU wastage bubble' [17] and observes that data sovereignty requirements, driven by the US CLOUD Act [3], are creating the first significant challenge to the dominance of US cloud hyperscalers in 15 years [5].

Present

Midha's company AMP has secured 1.3 gigawatts of compute infrastructure, financed with 80% debt [21, 30], and his incubation, Periodic Labs, is actively using LLMs and robotics to synthesize new materials in a 30,000 sq ft facility [8, 23].

Next 4 years

Midha posits that for Europe to achieve technological sovereignty, it must build a local compute infrastructure comparable in scale to Google's current 12-15 gigawatt capacity within this timeframe [10].

Suggested prompts

How does Anj Midha's 'at-cost' compute model for AMP create a sustainable business, and what are its primary competitive advantages against established hyperscalers? &nearr;What specific policy or market changes would be required to realize Midha's vision of a Western 'iron dome' for frontier model inference? &nearr;Beyond material science, what other scientific domains are most likely to experience the 'super-exponential gains' from compute that Midha describes? &nearr;To what extent is the 'misallocation of public capital' into traditional VCs, as claimed by Midha, a verifiable trend, and what are the implications for pension funds and endowments? &nearr;

Key concepts

AI Compute Infrastructure 9 ep Geopolitics & AI 5 ep AI for Science / Material Science 5 ep Venture Capital & AI Investment 4 ep Anthropic 4 ep AI Model Performance 4 ep Data Sovereignty 3 ep China's AI Strategy 3 ep Frontier AI Companies 3 ep Adversarial Distillation 2 ep

Notable quotes

“it's the first time in 15 years that the sort of hyperscaler dominance is up for grabs for startups.”

Anj Midha · The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha

“A primary reason for poor AI model performance in science is the lack of specialized physics and chemistry data on the public internet, which is mostly composed of text and code.”

Anj Midha · The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha

“we've started securing about 1.3 gigawatts of compute infrastructure that's roughly 40 billion dollars of cloud spend over the next four years”

Anj Midha · The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha

“I can tell you throwing more compute at the problem is probably having super exponential gains right now per iteration.”

Anj Midha · The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha

Report last updated: Apr 21, 2026

Get started free

Back to Entities Intelligence Report