Anthropic pays $500k–$850k for analog chip design engineers
The Seed Round: A Hardware Team Applying AI
Atrisa, a San Francisco startup founded in 2026, closed a $500,000 seed round in April as part of Y Combinator's Spring 2026 batch (Caplight reported). The company emerged from a rebrand (originally Refortif AI, the name still appears in the founders' email addresses) and sharpened its focus onto analog design after the founders watched the same bottleneck from inside the industry. YC group partner Nicolas Dessaigne backs the company.
The founding trio brings direct experience from the semiconductor trenches. Sayan Mitra, CEO, spent time at KLA Semiconductors as a research engineer and holds an electrical engineering degree from IIT Madras. Atman Kar, CTO, was previously an engineer at Texas Instruments, bringing low-level circuit design expertise to the technical architecture. Rithik Jain, COO, handles go-to-market strategy and customer experience, aiming to translate technical capability into enterprise sales. This is not an AI team learning hardware; it is a hardware team applying AI.
The product surfaces as a chat interface. An engineer describes a circuit function or a problem in plain English, and the AI agent parses the request, reasons about appropriate components and configurations, and can initiate simulations. The agent reasons hierarchically, breaking a system into blocks, matching topologies to specifications, and checking its own work against physical realities that trip up naive designs: reusability across older designs, possible interference, parasitics, and layout context. It plugs into the tools engineers already use, including process design kits (PDKs) and existing design documentation, so it augments a workflow rather than demanding a new one.
Atrisa's pitch is compact: describe a circuit in plain English, and the agent helps design it, debug it, and run the simulations. The company frames its product as "an assistant engineer right by your side" — not a replacement for the human, but a tireless second set of hands that has, in Atrisa's words, "internalized the intuition of an analog engineer." There is also a multi-agent angle. In a real design team, work stalls when everyone needs the one person who understands the bias network. Atrisa lets engineers hand pieces of a design to collaborating agents so blocks can move forward in parallel, without waiting on a single expert's calendar.
The differentiation lies in the proprietary dataset and heuristics the team is building to capture the nuances of analog design, a domain where general-purpose large language models typically falter. Right now, frontier LLMs struggle to make analog circuits. The knowledge lives in people, not documents. Atrisa's argument is that this knowledge can be encoded — and that whoever encodes it first owns a real moat. With $500,000 in accelerator funding from Y Combinator, the clock is now ticking toward a larger seed round. The milestones are clear: transition from a promising demo to a hardened, piloted product with at least one design partner, convert that pilot into a paid contract, and use that traction to secure a larger seed round, likely in the $3–5 million range (Startuply.vc reported), to scale the engineering team and formalize the go-to-market effort.
Why Analog Resists Automation
Digital design surrendered to automation decades ago. Abstraction layers, standardized cell libraries, and synthesis tools turned RTL into gates with minimal human intervention. Analog design never made that transition. The reasons are baked into physics: continuous signals, nonlinear device behavior, and performance metrics that pull in opposite directions across a high-dimensional space. A 2026 IEEE review of analog and mixed-signal design automation puts it plainly — unlike digital design, where scalable methodologies enabled extensive automation, analog design is characterized by tightly coupled decisions and nuanced trade-offs that resist abstraction.
Every analog block sits at the intersection of specification interpretation, topology selection, device sizing, physical layout, and simulation-driven refinement. Designers navigate these stages simultaneously because a choice in one ripples through the others. Change a transistor width to meet gain, and bandwidth shifts. Adjust bias current for noise, and power budget breaks. The 2026 IEEE paper notes that designers must navigate "tightly coupled decisions spanning specification interpretation, topology exploration, those stages, often under aggressive time-to-market pressure and growing system complexity." This coupling is why the design flow still hinges on manual reasoning and extensive simulation loops — a reliance that makes it vulnerable to slow turnaround, fragile design choices, aggressive overdesign, and a strong dependence on individual designer experience, intuition, and design style.
The expertise problem compounds the physics problem. Analog knowledge lives in heads, not repositories. A senior engineer knows which topology survives process variation at 28nm because she burned three tape-outs learning it. That knowledge transfers poorly. The 2025 IEEE study on emulating expert cognitive strategies identifies a persistent challenge: achieving generalizability — how insights from optimizing certain circuits can effectively inform the design of related circuits with differing topologies. Digital design solved this with standard cells and design rules. Analog has no equivalent. Each block is effectively a custom research project.
Automation attempts have repeatedly hit the same walls. Topology exploration suffers from time-consuming simulation-in-the-loop, reliance on expert-driven equation formulation, limited topological diversity, class-specific algorithm design, and risks to structural validity. Device sizing algorithms improve sample efficiency but remain largely restricted to predefined architectures. A 2025 arXiv survey of LLM-driven analog design automation concludes that "given the complex circuit-specific physics combined with the lack of standardized domain-specific training corpora, analog design automation applications remain in their early stages." Existing work stays in early stages; holistic joint optimization for practical end-to-end solutions remains largely unexplored.
The workflow itself reinforces the bottleneck. Topology exploration and circuit sizing run as separate sequential stages. When post-hoc parameter tuning cannot compensate for inherent topological limitations, the result is costly redesign cycles. After netlist generation, nominal parameters (transistor sizing, bias voltages, capacitances, resistances) only ensure functional correctness and often lead to suboptimal performance. Design flows remain heavily reliant on expert heuristics and extensive simulation cycles, limiting both efficiency and scalability.
Large language models cracked digital design first. They generate RTL, write TCL scripts for EDA flows, and autocomplete verification code because digital design speaks a formal, discrete language with abundant training data. Analog design speaks physics. The intricate interplay of circuit physics in deep submicron technologies and complex, multidimensional performance trade-offs has kept analog front-end design heavily dependent on expert intuition and iterative simulations, underscoring critical gaps in fully automated optimization for performance-critical applications.
Meanwhile, demand for analog content grows. AMS circuits quietly underpin IoT, edge computing, autonomous vehicles, 5G communication, and wireless sensing. AI training and inference require massive data movement between compute nodes, placing new demands on analog front-ends and mixed-signal IP. The global analog IC market is projected to grow from approximately $45 billion in 2023 to around $70 billion by 2030. The talent pool isn't keeping pace. U.S. semiconductor manufacturing faces a historic worker shortage as demand soars. The hand-crafted corner of chip design is also the one with the longest training curve — and the fewest new entrants willing to climb it.
Inside the Agent: From Prompt to Simulation
The interface is a chat window. Type a spec ("design a low-noise transimpedance amplifier for a 10 Gbps optical receiver, 50 Ω input, 3.3 V supply") and Atrisa's agent begins reasoning. It does not output a netlist in one shot. Instead, it decomposes the request hierarchically: system-level blocks first, then topology selection for each block, then sizing, then verification against the physical realities that separate a sim-only design from silicon that works.
This hierarchical reasoning is the core differentiator. The agent breaks a system into functional blocks (input stage, gain stage, output buffer, compensation network) and matches topologies to the spec constraints for each. It checks its own work against reusability (has this team solved a similar block before?), interference (substrate coupling, supply noise), parasitics (package, bondwire, on-chip routing), and layout context (matching, symmetry, guard rings). The company describes this as that internalized intuition — the undocumented rules of thumb that live in senior designers' heads and internal wikis, not in textbooks or public training data.
Integration with the existing toolchain is deliberate, not aspirational. The agent connects to proprietary process design kits (PDKs), design documents, and established analog design workflows. It plugs into the simulators engineers already run (Spectre, APS, Xyce) so the loop stays inside the environment the team has qualified. The workflow the company illustrates moves from "Serial and manual" (read constraints → edit schematic → run analysis → debug result, with repeated setup and tool switching) to "One continuous loop": set the design goal once, then design, simulate, debug, verify in a single thread while the engineer stays focused on judgment.
Simulation is integrated, not bolted on. The agent can invoke the simulator, parse results, and iterate, adjusting device sizes, tweaking compensation, re-running, without the engineer manually exporting netlists, configuring corners, or stitching together testbenches. The GitHub demo (a clean-room TypeScript/React prototype) shows this loop explicitly: editable spec intake fields with live constraint checks against mock simulation runs, a topology/library explorer with agent-ranked synthetic candidates for a TIA, an instrumentation amplifier, and a bandgap reference, and simulation sandbox scorecards using deterministic seed data. Agent suggestion cards feed a human review queue; mock EDA integration status cards signal where the real PDK hooks would land.
Layout context is acknowledged but deferred. The roadmap reads: reasoning today, layout next, post-silicon debug after. The current agent reasons about parasitics and layout constraints but does not yet generate GDS. That step, when it arrives, will need the same hierarchical awareness: floorplan-aware parasitic extraction, IR-drop analysis, and the layout-dependent effects that make analog layout a discipline of its own.
A multi-agent layer sits on top. This enables the parallel workflow described earlier. Each agent carries context (the spec, the PDK corners, the team's design rules, the history of prior iterations) and surfaces its reasoning for human review before committing changes.
The agent readiness score from independent tracking sits at 10/100, labeled "Early." No public agent surfaces have been detected yet. The product is enterprise SaaS, priced on contact, wired into proprietary PDKs and workflows. But the architecture (hierarchical reasoning, toolchain-native simulation, multi-agent parallelism, layout-aware constraints) maps directly to how analog design actually gets done, not how a generic LLM hallucinates a schematic.
The Engineer's New Loop: Arbiter, Not Operator
The analog engineer's day job has always been a cycle of simulate, tweak, simulate again — hours lost to parameter sweeps that no script quite captures because the intuition guiding them lives in a senior designer's head, not a version-controlled repo. Atrisa's agent and its research cousins (AnalogAgent, AnaFlow, RFChipAgent) don't just accelerate that loop; they invert it. The engineer stops writing SPICE decks by hand and starts reviewing the agent's proposed sizing updates, flagged spec violations, and Chain-of-Thought rationale, then approves, rejects, or edits the JSON-structured suggestion. The human becomes the arbiter of intent, not the operator of the tool.
That shift only works if the agent learns what the senior designer knows. AnalogAgent's Knowledge Curator distills each optimization cycle (simulation feedback, constraint checks, the reason a 15% width increase fixed phase margin without blowing the power budget) into an Adaptive Design Playbook stored in a self-evolving memory. The playbook then feeds targeted guidance back to the Code Generator, so the next design starts with last week's hard-won heuristic instead of a blank prompt. In benchmarks, this structure lifted compact models like Qwen-8B by nearly 49% average Pass@1 across tasks, reaching 72% overall — evidence that structured memory compensates for smaller parameter counts, a practical necessity when proprietary PDKs forbid cloud-hosted LLMs.
The collaborative loop is explicit: human enters netlist and target specs; agent decomposes tasks, schedules simulations, proposes constrained parameter updates (typically 20% or less per step), and presents reasoning traces with symbolic calculations (gm ∝ (W/L)(VGS−Vth) for gain-bandwidth trade-offs). The designer sees flagged violations (a transistor drifting into subthreshold, a symmetry mismatch) and decides whether to accept the fix, relax a spec, or override the topology. AnaFlow demonstrated this on a folded-cascode OTA: 64 total simulations versus more than 1,000 for reinforcement-learning baselines, with full spec compliance and stepwise audit trails the human could interrupt at any reasoning stage.
But the research surfaces a trap. A 2024 MIT review of over 100 human-AI studies found that on average, human-AI combinations underperform the best AI-only system — statistically significantly worse. The combination wins only on creation tasks (bird classification: human 81%, AI 73%, together 90%) and loses on decision-making tasks. Analog design straddles both: topology selection is creative; sign-off is a decision. The danger is false reassurance. As UCSF's Robert Wachter said, AI is "right often enough to be useful and wrong often enough not to be entirely trusted." If the agent nails consecutive sizing runs, the human stops scrutinizing the next one. That vigilance decay is a known human failure mode, not a model flaw.
Skill atrophy compounds it. Stanford researchers coined "never-skilling" — trainees who outsource cognitive reps to AI never build the independent reasoning to catch the next error. In analog, the "reps" are the manual iterations that teach a designer why a cascode mirror needs that specific bias voltage. If the agent always proposes the fix, the junior engineer never internalizes the physics. The research warns that edge cases (topology-level changes, bias network redesign, strict area constraints) still demand human intervention. An engineer who never wrestled with those cases won't recognize when the agent hallucinates a plausible-looking but non-functional topology.
Governance must move from ad-hoc to embedded. SmartBrief's framework argues leaders should define three boundaries before deployment: what the agent does autonomously, what it recommends for human approval, and what it never touches. Permissions, escalation paths, audit trails, confidence thresholds, and shutoff controls become part of the design flow, not a PDF reviewed quarterly. The endgame is shifting from "human in the loop" — approving every sizing step, to "human above the loop": setting guardrails, managing exceptions, interpreting context, redesigning the workflow itself. That transition is where the talent market will fracture.
Incumbents Strike Back: The Agentic EDA Arms Race
The three EDA giants (Cadence, Synopsys, and Siemens) have spent the past year rewriting their roadmaps around a single word: agentic. At DAC 2026, each unveiled a flagship "super agent" or agent platform, and the messaging converged on the same claim: EDA is moving from tool-by-tool workflows to goal-driven, multi-agent execution. For a three-person YC startup like Atrisa, the incumbents' moves define the competitive terrain more than any investor memo could.
Cadence's strategy is the most visible and the most layered. Virtuoso AI Studio, the latest evolution of a platform with nearly four decades of analog lineage, now sits atop a four-stack agent architecture. ChipStack AI Super Agent targets front-end digital design and verification (RTL generation, testbench creation, verification planning, regression orchestration, debugging) and runs on NVIDIA's Nemotron 3 Ultra model. InnoStack AI Super Agent handles digital implementation and signoff. AuraStack AI Super Agent extends into PCB and advanced packaging with a multiphysics foundation that concurrently models electrical, thermal, and mechanical behavior. And ViraStack AI Super Agent, the one that matters for analog, sits inside Virtuoso Studio to help with schematic creation, circuit optimization, testbench development, and layout migration. Cadence's own white paper frames this as a five-level maturity ladder: Level 1 (ML-powered optimization), Level 2 (conversational AI), Level 3 (task-level agents), Level 4 (coordinated multi-agent workflows), and Level 5 (full autonomy). At DAC, the company showcased agentic AI across the design continuum.
Synopsys took a different architectural bet. AgentEngineer™ sits on the Synopsys Autopilot™ Platform, described as an open, secure foundation for autonomous engineering. The emphasis is on token efficiency and latency reduction across complete workflows rather than a portfolio of domain-specific super agents. Siemens, meanwhile, positioned Fuse EDA AI Agent as an end-to-end orchestrator spanning architectural exploration, RTL coding, digital and custom IC design, verification, place-and-route, physical sign-off, and manufacturing readiness. Its partnership with NVIDIA targets "self-verifying, long-running EDA AI agents", a direct answer to the trust gap that keeps engineers from handing off signoff-critical steps.
All three incumbents share two advantages Atrisa cannot match: installed base and process-node certification. Cadence's flows are certified for TSMC N2P, A16, and A14; Samsung Foundry's second-generation 2nm; and the company has demonstrated a 60 percent faster design closure with 10X agentic workflow productivity on a 6 GHz analog/RF PHY migration. Synopsys and Siemens each count every major foundry and IDM as a reference customer. When a design team at a hyperscaler or automotive Tier 1 evaluates an AI agent, the incumbent's integration with their existing signoff stack, DRC/LVS decks, and IP libraries is the default — not a proof-of-concept.
Where Atrisa differs is scope and entry point. The incumbents build agents that live inside their own ecosystems; Atrisa's agent is designed to work with the simulators and design tools engineers already use, translating natural-language intent into circuit topology, sizing, and simulation setup without requiring a platform migration. Its founding team (ex-KLA and Texas Instruments engineers) built the agent around the analog designer's mental model: hierarchical reasoning about topologies, parasitics, layout context, and reuse across older designs. That specificity is a wedge. Cadence's ViraStack operates inside Virtuoso; Synopsys's agents assume a Synopsys-centric flow. Atrisa targets the engineer who writes a prompt and expects a simulated schematic back, not a consulting engagement to onboard a new platform.
The incumbents have noticed. Cadence's white paper explicitly states that AI "is NOT a replacement for the analog engineer" but "an assistant to and, in some cases, an extension of the engineer." That framing leaves room for a specialized agent that plugs into existing workflows rather than replacing them. Synopsys's Autopilot platform is marketed as open; Siemens emphasizes Fuse's customization flexibility. In practice, "open" in EDA has historically meant "open to our partners' IP." But the pressure to demonstrate interoperability is real, NVIDIA uses Cadence solutions internally while also collaborating on Nemotron, and the foundries themselves are pushing "agent-ready" flows that could create integration points for niche agents.
The arms race is also a talent race. Atrisa's three founders represent a different hiring profile: analog designers who learned to build agents, not AI researchers learning analog. That distinction matters when the product's core claim is that it "thinks like an analog engineer."
The next signal to watch isn't a press release. It's whether a lead customer at a major design house runs Atrisa's agent alongside Virtuoso or Synopsys Custom Designer for a real tapeout block, and whether the incumbent's account team treats that as a complementary evaluation or a competitive displacement. The incumbents have the platform; Atrisa has the wedge. The market will decide if the wedge becomes a crack.
Talent: The Analog Brain Drain and New Skill Sets
The semiconductor workforce crisis has a specific analog face. Forty-four percent of open roles sit in analog design, another 43 percent in embedded engineering, and the pool of qualified candidates is shrinking while chip demand surges across AI, EVs, 5G, and IoT. U.S. chipmakers — Samsung and Micron among them, have sounded the alarm on a historic labor shortage that predates the current AI boom but has been sharpened by it.
What began as competition for skilled engineers has hardened into a systematic talent drain reshaping the startup ecosystem. Elite engineers who moved AI systems from experimentation into production are now choosing heavily funded private companies and AI-native startups over traditional EDA incumbents and big-tech semiconductor divisions. The shift is visible in compensation: generative AI managers in finance command a median base salary of roughly $190,000, while Anthropic lists a Research Engineer in Chip Design RL at $500,000–$850,000. Traditional analog roles rarely approach those bands.
The analog talent pipeline was already narrow. Across India and globally, practicing analog circuit designers typically hold at least an MTech or PhD — a credential floor that limits supply. New AI PhDs in the U.S. and Canada rose 22 percent from 2022 to 2024, but the incremental graduates took academic positions, not industry jobs. Meanwhile, the flow of AI researchers and developers into the United States has collapsed 89 percent since 2017, with an 80 percent drop in the last year alone.
Agents are rewriting the job description. Job postings referencing "agent orchestration" exploded 1,721 percent year over year, per Draup data shared with CNBC; LangGraph mentions jumped 679 percent, LlamaIndex 291 percent, and RAG 259 percent. Governance-related skills (responsible AI, AI risk management) now outnumber model-training references nearly two to one. For analog engineers, the new stack layers agentic workflows (LangGraph, CrewAI), retrieval-augmented grounding, evaluation harnesses, and deployment observability on top of SPICE, layout parasitics, and device physics.
Incumbents are responding with internal reskilling. Major banks (and by extension, large chip design houses) are training existing domain experts to become "forward-deployed engineers" who integrate agents into verification, migration, and sign-off flows. The priority, hiring leads say, is soft skills paired with technical depth: problem solving, the courage to ask tough questions, and the ability to decide when a human must stay in the loop.
The salary premium for hybrid profiles is real. Databricks shows a board median of $250,000 across 491 salaried roles; Anthropic sits at $385,000 median across 562 roles. But the analog core (device intuition, noise analysis, stability across corners) remains resistant to full automation. An analog engineer at HCL said the core cannot be grabbed by AI; layout acceleration and technique suggestion are the near-term ceiling, and even that may take 15 years.
Hiring managers now face a two-track problem: find the rare engineer who speaks both analog and agentic, or build the bridge internally. The startups moving fastest (Atrisa among them) are betting on the latter, embedding the tribal knowledge of ex-KLA and TI veterans into agents that a new generation can steer. The winners will be the teams that treat AI literacy not as a bonus skill but as table stakes for every analog seat.
Market Outlook: A $1.45 Billion Curve
The AI-assisted analog circuit design market is tiny today but growing faster than the broader EDA stack. Semiconductor Insight pegs it at $0.68 billion in 2025, rising to $1.45 billion by 2034 — a 9 % CAGR. A narrower slice, the AI analog circuit sizing automation engine market, started at $215 million in 2025 and is forecast to hit $540 million by 2034 (10.2 % CAGR). For context, the entire EDA software market was $14.55 billion in 2025 (Gartner's data shows) and tracks toward $34.71 billion by 2035 (9.08 % CAGR), while the analog AI chip market (hardware that runs inference in the analog domain) is projected to explode from $250.85 million in 2025 to $2.45 billion by 2035 (25.6 % CAGR). The analog circuits market itself, the ultimate addressable space, sits at $17.13 billion in 2025 heading to $27.59 billion by 2033 (6.1 % CAGR).
| Market Segment | 2025 Size | 2034/35 Forecast | CAGR | Source |
|---|---|---|---|---|
| AI‑Assisted Analog Circuit Design | $0.68 B | $1.45 B (2034) | ~9 % | Semiconductor Insight (2026‑07‑14) |
| AI Analog Circuit Sizing Automation Engine | $215 M | $540 M (2034) | 10.2 % | Semiconductor Insight (2026‑07‑03) |
| EDA Software (total) | $14.55 B | $34.71 B (2035) | 9.08 % | Precedence Research (2026‑02‑02) |
| Analog AI Chip | $250.85 M | $2.45 B (2035) | 25.6 % | Precedence Research (2026‑09‑15) |
| Analog Circuits (total) | $17.13 B | $27.59 B (2033) | 6.1 % | Verified Market Reports (2026‑05‑24) |
North America dominates the AI-assisted analog design segment because the major EDA vendors (Cadence, Synopsys, Ansys, Siemens EDA, Keysight) are headquartered there and have embedded AI modules (Cadence Cerebrus, Synopsys Custom Compiler) into their flagship suites. Asia‑Pacific shows the fastest growth rate, driven by foundry expansions in Taiwan, Korea, and China and by government-backed semiconductor self‑sufficiency programs.
The adoption curve is being pulled by three forces. First, circuit complexity in automotive electrification, IoT edge devices, and 5G RF front‑ends has outstripped the supply of seasoned analog engineers. Second, time‑to‑market pressure is measurable: design teams using AI‑driven synthesis tools report compressing the analog layout phase from weeks to days, cutting design‑verification loops and moving from concept to silicon in weeks rather than months. Third, semiconductor equipment manufacturers are investing in AI‑compatible EDA platforms, lowering the cost of entry for mid‑size firms.
"AI tools are no longer optional; they are becoming the baseline for achieving design robustness in a fragmented component ecosystem." — Semiconductor Insight (2026‑07‑14)
Watch these adoption signals over the next 12‑18 months:
- Specialized vertical AI engines for power‑management, RF front‑ends, and sensor interfaces that reduce the need for broad‑scope training and enable faster deployment in niches like IoT wearables or autonomous‑vehicle subsystems.
- Foundry PDK‑integrated AI workflows where process design kits are embedded directly into the AI agent, cutting back‑end iteration cycles and creating new revenue streams for both EDA vendors and foundries.
- Open‑source analog design datasets maturing enough to let startups train cost‑effective models, challenging the premium integrated suites from incumbents.
- Edge AI deployment acceleration: the broader edge AI market is growing from $18 billion to $60 billion by 2030 (41 billion in 2025, 75 % YoY), and Gartner projects two‑thirds of enterprises will deploy edge AI by 2029 versus 10 % in 2025. That pull creates demand for ultra‑low‑power analog front‑ends that only AI‑assisted sizing can deliver on schedule.
- Neuromorphic and in‑memory compute silicon shipping today (Innatera, BrainChip) promising 100× power efficiency gains, these chips need analog design flows that can handle spiking neural network topologies, a natural fit for agentic tools.
The main friction points remain: a talent gap in ML workflows among analog houses, proprietary and inconsistently documented design data that limits model fidelity, automotive/medical certification overhead for AI‑generated schematics, IP‑ownership concerns with cloud‑based training, and high upfront licensing fees that keep early‑stage startups on the sidelines. The inflection point arrives when the cost of not adopting (measured in respins, delayed tape‑outs, and engineer burnout) exceeds the license fee. For the analog design community, that crossover is approaching fast.
At Atrisa's office, the GitHub demo still runs on mock data. The $500,000 buys runway. But the wedge is real: a chat window that speaks the language of parasitics, matching, and bias networks — because the people who built it learned that language the hard way, at the bench, before they ever wrote a line of agent code.
Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic, and the people building the field.