The $10 Query: The Compounding Cost of AI Hallucinations

This first installment of our DVCon US ’26 video series dives straight into a reality check that often gets lost in the hype of generative AI: The cost of a single "thought."

In this segment, Srini (AsFigo) sits down with Asif ETV (HPCINFRA) to discuss why the transition to AI-native chip design isn't just a software challenge—it’s a massive infrastructure and economic hurdle.

Watch the Full Segment: In this 3-minute clip, watch Asif and Srini break down the economic reality of the modern AI-EDA stack and why optimizing for the right infrastructure is the only way to stay within budget while bringing AI into production.

The Cost Explosion: Doing the Math on Hallucinations

When we talk about "Agentic AI" in EDA, we imagine autonomous loops of design and verification. But as Asif ETV points out, the meter is always running, and it’s running faster than many teams realize. Sharing a startling metric from recent infrastructure experiments, the math of using high-end models for design queries becomes clear very quickly:

"It was like, just one query, $10. That’s all. Now, whether we gave the correct result or not, we don’t know. Now, multiply that... how many queries in a day one engineer would make times how many analytical engineers?"

The "neutrality of the economic slide" is a compounding problem. $10 is negligible for a one-off task, but Agentic AI is built on iteration. When an agent requires 20, 50, or 100 queries to code a module, and you multiply that across an entire engineering team, you aren't just looking at an expensive tool—you're looking at a capital burn rate that can reach thousands of dollars per day before a single gate is even synthesized.

If those queries result in hallucinations, you haven’t just lost time—you’ve burnt a significant hole in your budget on a mistake. For a startup, this makes "unprotected" AI deployment a financial non-starter.

Infrastructure: The Silent Security Risk

Beyond the dollar sign, there is the critical question of IP security. Asif noted that while the industry has spent 40 years hardening traditional EDA environments, AI infrastructure is often treated as an afterthought by teams moving quickly to catch the wave. You cannot scale AI if your infrastructure is a "leaky bucket" for either your IP or your budget.

The AsFigo Perspective: Bridging the Gap

This is where the AsFigo philosophy provides a solution. If the inference itself is expensive, the verification of that inference must be nearly free. As Srini noted, this is why the "EDA license cards" and open-source movements are vital to the stack.

AsFigo’s values are built on this bridge:

  • Economic Viability: Using open-source "Utility" tools (Verilator, SVALint) to create a zero-marginal-cost sandbox. This allows AI agents to fail, learn, and correct thousands of times without incurring additional license or inference "taxes."
  • Security & Stability: Moving AI from an "ad-hoc" experiment to a professional-grade, secure floor where IP is protected.
  • Democratization: Ensuring smaller firms aren't priced out of the AI revolution by the combined weight of exponential inference costs and commercial license gates.

Watch the Full Segment: In this 3-minute clip, watch Asif and Srini break down the economic reality of the modern AI-EDA stack and why optimizing for the right infrastructure is the only way to stay within budget while bringing AI into production.

https://youtu.be/Ir6Xn8M2y2w 

Stay tuned for our next installment, where we look at how open-source front-ends are being used to provide real-time PPA feedback to AI agents.

Comments

Post a Comment

Popular posts from this blog

Join Us in Cambridge - Advancing Chip Verification with UVMLint & SVALint

AsFigo invites you to our next technical session, focused on enabling semiconductor design teams to adopt open-source EDA tools—especially for testbench linting workflows. Remote-dial-in link:  UVMLint meetup - Cambridge, June 27 | Meeting-Join | Microsoft Teams 📍 Event: UVMLint & SVALint – Enablers to Move Your Chip Design Ahead 📅 Date: Friday, June 27, 2025 🕙 Time: 10:00 AM to 12:00 Noon (UK Time) 📌 Location: Wellington House, East Road, Cambridge, CB1 1BH, United Kingdom ☎ Phone: +44 1223 451000 🌐 Format: Hybrid (in-person & remote via Microsoft Teams) 💡 Cost: Free (registration required - https://forms.gle/upzLGyWJbRnbgRzU7 ) After a successful SVALint session in Reading, UK, we’re bringing this discussion to Cambridge—one of the UK’s most active semiconductor hubs. This event is geared toward engineers working in UVM based verification, formal verification, and testbench architecture using UVM and SVA. Ag...
Read more

SVALint Technical Meetup – Reading, UK

SVALint Technical Meetup – Hybrid Event in Reading, UK AsFigo invites engineers and verification professionals to a focused technical meetup on SVALint —an open-source static verification abstraction linting framework. The event takes place on Sunday, 8 June 2025 , from 15:00 to 18:00 BST , hosted in a hybrid format (face-to-face and online). Register via:  https://forms.gle/qo8TBczXjNk62yAP6   Venue (In-Person) Meeting Room-5, First floor,  Novotel Reading Centre 25B Friar Street – RG1 1DP, Reading – United Kingdom T +44 (0) 1189 522 610 novotel.com/5432 Virtual Attendance Join via Microsoft Teams Why SVA Linting? SystemVerilog Assertions (SVA) play a critical role in ensuring correctness in complex verification environments. However, poorly structured assertions can hurt simulation performance, complicate debugging, or lead to ambiguous semantics. Linting SVAs helps catch such issues early—enforci...
Read more

Cracking the UVM-Verilator Code: 50+ IPs, AI Guardrails, and the Open-Source North Star

UVM on Verilator Isn’t Impossible. It’s Just Hard. We Did It Anyway. The “standard” line in the industry: UVM and Verilator don’t mix. Enterprise-grade open-source verification is a fantasy. So we put that claim under load. ✔ 50+ production UVCs ported ✔ 35 + non-obvious failure modes uncovered ✔ Converted into UVMLint and SVALint rules ✔ AI guardrails to prevent regressions This post is the why and what; the how and numbers are for the room at Hacker Dojo.  This is production-grade code—the kind that tapes out chips—running on open-source infrastructure. The implication is bigger than a port: Reduced dependency on closed ecosystems. Negotiating leverage with EDA vendors. Strategic control over your verification stack. If you’re a semiconductor company exploring Verilator for serious IP, don’t start from scratch. Start from the team that already mapped the minefield. 📅 Join us at the Hacker Dojo Next Thursday, the AsFigo team...
Read more