India's USD $200B AI hub & Claude builds C compiler
Experts from IBM discuss Google's $200B AI investment in India, Claude's autonomous C compiler creation, the significant security risks in AI agent skills, and the looming AI ROI problem facing IT leaders, debating the shift from per-token to value-based pricing.
Google’s $200B AI Infrastructure Deal in India
The discussion begins with Google and DeepMind’s massive potential investment of up to $200 billion to build a comprehensive AI hub in India. The experts debate the primary motivations behind this historic deal.
Kush Varshney frames it as a geographical and geopolitical shift, suggesting it’s about recreating a “golden road” where India becomes a central hub for technology and culture, connecting to South Africa, Australia, and Singapore. He sees it as India positioning itself as a leader for the “global majority” and the “next billion” users.
Mihai Criveti offers a more skeptical perspective, questioning if the sum is truly “staggering” when contextualized against the cost of tokens at scale or the price of a new hyperscaler data center. He posits that the main driver isn’t access to cheap power but rather data sovereignty. > “We’re seeing a lot of countries and organizations try to say we need to have the AI in our country or in our region or in our geography… So we don’t have a dependency on the US or we don’t depend on data centers or access to these models. They can always be taken away if there is a geopolitical issue.”
The consensus is that the investment is less about simple economics and more about establishing regional control and sovereignty over critical AI infrastructure.
Claude Autonomously Builds a C Compiler
The conversation shifts to a notable achievement from an Anthropic researcher who used 16 AI agents to autonomously build a 100,000-line C compiler.
Mihai Criveti expresses surprise that this “hasn’t been done earlier,” stating that serious practitioners have been using AI in continuous, iterative loops for years. The key elements of this advanced approach include:
- Continuous Mode: Generating code, running unit tests, performing code reviews, and regenerating in a loop until the results are satisfactory.
- Multi-Model Strategy: Using different models for tasks they excel at, such as Claude for fast code generation and other models for robust code review.
- Tooling is Essential: AI agents need the same tools as human developers—IDEs, linters, debuggers, and deployment environments. Simply giving an agent a blank slate is ineffective. “AI helps automate some of the coding aspects of software development. It doesn’t replace the art of software engineering.”
The panel agrees that while impressive, the human role remains critical. Martin Keen notes the researcher still had to step in to redirect the agents’ efforts. However, he highlights the dramatic efficiency gains, comparing the $20,000 in API costs and two-week timeline to a COCOMO model estimate of $2.5 million and 20 months for a human team.
This raises concerns about the future, with Martin echoing the researcher’s “unsettled feeling that the moat of human requirement is getting smaller and smaller.” Mihai warns of a potential expertise gap: as agents automate junior-level tasks, it will become harder for new developers to gain the foundational experience needed to become the senior experts required to guide these complex agentic systems.
Security Vulnerabilities in AI Agent Skills
A recent Snyk report revealing that 36% of AI agent skills contain security flaws serves as a sobering reality check. The vulnerabilities range from unintentional flaws, like hardcoded test tokens, to actively malicious payloads and prompt injections.
Kush Varshney is “very, very concerned,” viewing this as a societal and governance issue for skill marketplaces, not just a technical one. The panel highlights the unique danger of AI agents due to their autonomy and the broad access they are often granted, creating a potentially massive blast radius from a single compromised skill.
The discussion on solutions centers on governance and control:
- Technical Guardrails: Mihai Criveti describes the extensive work on
context-forge-gatewayto secure interactions using PI filters, sandboxing (gVisor, Firecracker), scanners, and trusted catalogs. However, he stresses that technology alone is insufficient. - Enterprise Governance: For corporations, a locked-down, internal “app store” for vetted skills is essential, much like enterprises prevent sideloading on corporate iPhones.
- Individual Responsibility: For individual users, the advice is to experiment in isolated environments like a VM or an old laptop with no sensitive data. Mihai’s core principle is to “only give your agents access to the things you’re comfortable with it messing up.”
- The Marketplace Dilemma: Martin Keen points out the tension between security and openness. A heavily regulated, App Store-style review process for skills would likely be unpopular and could stifle the open ecosystem that makes agents so powerful.
The Inevitable AI ROI Problem
The final topic addresses the growing concern among IT leaders about the return on investment (ROI) from AI. An Apptio report found that 9 in 10 IT leaders say doubts about value are affecting their AI investment decisions.
Martin Keen calls this “the most predictable thing ever in business.” The initial phase of hype-driven, open-checkbook investment is naturally followed by a period of scrutiny where leadership asks, “Is this actually doing anything useful for us?”
The core challenge is that the ROI of AI is difficult to quantify. In response, the experts predict a fundamental shift in pricing models. Both Kush Varshney and Mihai Criveti believe the industry will move from per-token pricing to value-based pricing. > “I’m just thinking how long before we’re going to see the charging model from a lot of these providers… switch from metering and pricing per token to metering and pricing per outcome. Because that’s when everything is going to explode.” - Mihai Criveti
Mihai notes that his conversations with clients have already shifted from “how to use AI” to “how much does it cost?” He argues that organizations cannot even begin to measure value until they solve the governance problem of “random acts of AI” and “shadow AI” across departments, which makes it impossible to even accurately track costs.