Computing at the Crossroads
In her ISCA 2026 keynote, our CEO, Dr. Debbie Marr, explains why now is the time to build a high-performance, general-purpose RISC-V CPU and which shifts in architecture, economics, and software make it possible.
In her ISCA 2026 keynote, our CEO, Dr. Debbie Marr, explains why now is the time to build a high-performance, general-purpose RISC-V CPU and which shifts in architecture, economics, and software make it possible.
We are designing a high-performance, general-purpose RISC-V CPU. Our focus is per-core performance, because it sets the pace for AI infrastructure, the data center, and the devices you use every day. A core with higher per-core performance can also match a slower core at a lower voltage and frequency, which reduces power.
A better computer puts its owner's interests ahead of any third party that profits from the owner's data or dependence. It protects the owner's money, time, and attention, and it does what the owner tells it to do. We design our CPUs for computers like that.
Accelerators perform the parallel mathematics of AI. The CPU runs everything around it, including data preparation, scheduling, networking, and the control logic of agentic workloads. When the CPU falls behind, the accelerators wait.
FIG. 03 CPU-to-GPU ratio in AI systemsThe ratios come from Lip-Bu Tan, chief executive officer of Intel, in April 2026.
“The most important thing is single-threaded performance... If the CPU gets throttled, then we're holding back a whole bunch of GPUs.”Jensen Huang, chief executive officer of NVIDIA, GTC 2026, March 2026
“CPUs are making a massive comeback: 35% annual growth.”Lisa Su, chief executive officer of AMD, May 2026
“Earlier predictions of a 4x increase in CPU core requirements are now being discussed at 8x or even 10x by the industry.”Rene Haas, chief executive officer of Arm, June 2026
Amdahl's Law sets a limit. The serial portion of a workload caps the speedup that parallel hardware can deliver, so each added core yields a smaller gain. Raising per-core performance raises the speedup at every core count.
At advanced process nodes, every square millimeter of silicon is expensive. For any workload with a serial fraction, higher per-core performance raises throughput at every core count, and system throughput is what you pay for.
Adjust the controls to test the relationship.
Leading-edge wafers take longer to fabricate and cost more at every node. Advanced packaging adds months of lead time and significant cost per unit. Under these conditions, hardware returns depend on longevity and time to market. Specialized accelerators face a harder path. General-purpose processors win on workload diversity, longevity, and time to market.
Hyperscale cloud providers, including Microsoft, have spent the past decade moving workloads away from x86. The toolchains, memory models, and migration practices built for that transition now apply directly to RISC-V, and GCC, LLVM, and the Linux kernel treat RISC architectures as first-class targets.
Intel positioned the Pentium Pro, its first P6-family processor, for servers and workstations. It disproved the belief that x86 could not be fast. Dr. Debbie Marr served as its server architect.
Virtually all workloads ran on Intel chips until we invented Graviton in 2018.
Andy Jassy, chief executive officer of Amazon, made this statement.
Arm-based Windows PCs reach the mainstream with Qualcomm. NVIDIA announced its N1X PC processor at Computex 2026.
Google migrated more than 30,000 applications from x86 to Arm and published a path to datacenter-scale RISC-V.
The SHD Group projects that RISC-V will reach 33.7% of processor revenue in 2031, up from 2.5% in 2021.
FIG. 06 Milestones in the architecture transitionThe 2031 value is a projection by The SHD Group. NVIDIA's N1X announcement was reported by Entrepreneur in June 2026. Google reported the 30,000 applications in “Using AI and automation to migrate between instruction sets,” October 21, 2025.
In 1996, the industry said x86 could not be fast. Today it says the same about RISC-V.
The same toolchains and practices that carried hyperscale software from x86 to Arm now lead to RISC-V.
We combine the openness and extensibility of RISC-V with a new microarchitecture designed to show that an open instruction set can lead on performance and power efficiency in data center, client, mobile, and embedded markets.
RISC-V carries no x86 legacy and none of Arm's licensing costs and constraints. We spend our design effort on performance rather than on compatibility.
Our core is wide, deep, and out of order, and we design it to hide memory latency and find instruction-level parallelism at low energy per instruction. Per-core performance is the key to multi-core performance.
The architects behind Intel's Pentium Pro, Nehalem, and Haswell generations are designing it, and they build in the reliability that data centers expect from the start.
AheadComputing has raised $53M to date. The $30M Seed2 round was co-led by Eclipse, Toyota Ventures, and Cambium.
Eclipse is a venture firm that partners with founders transforming the physical economy and its critical industries.
eclipse.capital ↗Toyota Ventures is Toyota's early-stage venture capital arm and invests in frontier and climate technologies.
toyota.ventures ↗Cambium Capital is an early-stage venture capital firm that invests in advanced compute, deep tech, and AI.
cambium.vc ↗Corner Ventures backs high-growth technology companies across the US, the Middle East, and Asia.
corner.xyz ↗Trousdale Ventures backs space, mobility, national resilience, deep tech, and consumer companies.
trousdale.vc ↗EPIQ Capital Group is an investment network that gives founders and families institutional-grade access to private investments.
epiqcg.com ↗MESH is an electronics-focused venture capital firm with relationships across the industry, from foundries to ODMs.
mesh.vc ↗Stata Capital Partners is a deep tech investor, founded by Analog Devices founder Ray Stata, that backs early-stage founders.
statacapital.com ↗Our first product is in development through collaborations with semiconductor ecosystem partners, alongside design exploration with hyperscalers and compute providers.
Alchip Technologies is a Taipei-headquartered provider of high-performance ASIC design and production services.
alchip.com ↗Cadence is a computational software company, headquartered in San Jose, California, that enables electronic and system design.
cadence.com ↗SkyeChip is a Malaysia-based IP and IC design company that serves AI and high-performance computing.
skyechip.com ↗Tenstorrent is a next-generation computing company that builds computers for AI and has deep RISC-V expertise.
tenstorrent.com ↗Company descriptions are drawn from each organization's public materials. Logos are trademarks of their respective owners.
Agentic AI is creating urgency. CPUs bottleneck accelerator throughput, and agentic workloads run largely on CPUs.
The software ecosystem is in motion. RISC-V is a clean-slate architecture, and AI-enabled toolchains ease the transition.
Manufacturing time and cost are rising. General-purpose designs win on workload diversity, longevity, and time to market.
We are hiring CPU architects, performance modelers, and verification engineers who want to define high-performance computing for the next decade.
Three industry shifts are converging: AI systems are becoming CPU-bound, the software ecosystem is moving beyond x86, and silicon economics now reward general-purpose design. We founded AheadComputing to meet all three.
Every real workload alternates between phases that spread across many cores and phases that run on one. During the serial phases, additional cores sit idle. A faster core shortens every phase, which is why per-core performance is the key to multi-core performance.
We chose RISC-V because it lets us license our designs and sell our own silicon. We start by licensing CPU Core IP and Compute Core Complex IP to semiconductor companies, directly and through ASIC design service partners. We then build on the same cores to offer a compute chiplet in a system-in-package. RISC-V is an open specification that anyone may implement without paying for an instruction set license, and that freedom carries through to our licensees.
The specification is open, and our implementation is proprietary. We develop commercial technology that conforms to the RISC-V specification, so software built for the RISC-V profiles we support runs on our cores, and the microarchitecture that makes those cores fast belongs to AheadComputing. RISC-V also separates a small, stable base from standardized extensions, so our design effort goes to execution rather than to decoding decades of accumulated encodings.
AI is speeding up software development. The Jevons paradox is the observation that making a resource cheaper to use tends to increase total use of it. We expect it to apply to software. As productivity rises, the volume of software will grow, new categories of software will become practical, and the pool of developers will expand.
Meanwhile, wafer fabrication and advanced packaging take longer and cost more. By the time specialized hardware ships, the workload it was designed for may have moved on. General-purpose processors serve the widest range of workloads over the longest service life.
Processor performance comes down to one equation, the Iron Law of Performance. Our product details remain confidential, so these principles describe the direction of our work.
We also weigh the energy each instruction costs.
RISC-V's standard extensions, such as bit manipulation and vectors, let compilers express the same work in fewer instructions.
Instructions per programKeep the core productive while memory responds, through deep out-of-order execution and aggressive speculation.
Cycles per instructionFind independent work and execute it in the same cycle, so single-thread performance scales with the workload.
Cycles per instructionIn superchips that integrate CPU, GPU, and accelerators, compute performance per mm² is the defining metric. We optimize power, performance, and area together, so each cycle is as short and as efficient as its silicon allows.
Time per cycle · energyFIG. 15 Four design principles, each tied to a term of the Iron LawThis conceptual model is not an AheadComputing design.
Dr. Marr closed her ISCA 2026 keynote with a call to the research community. These are the areas where she sees the greatest opportunity for new ideas.
We design our CPU cores for workloads where per-core performance and power efficiency determine the result. Our markets span data centers and AI, workstations and PCs, phones and tablets, and embedded and edge systems.
McKinsey projects that by 2030 about 70% of demand for data center capacity will be for facilities equipped to host advanced AI workloads. Within that demand, McKinsey projects that inference will overtake training by 2030 and account for more than half of AI compute. Arm estimates that as agentic applications scale, data centers will need more than four times their current CPU capacity per gigawatt. Inside every AI server, the host CPU prepares data, schedules work, manages memory, and moves results across the network. When it falls behind, accelerators that cost far more than the CPU sit idle.
The 70% projection is the midpoint of McKinsey & Company's scenario range, with demand measured in power, from “AI power: Expanding data center capacity to meet growing demand,” October 29, 2024. The inference projection comes from McKinsey & Company, “The next big shifts in AI workloads and hyperscaler strategies,” December 17, 2025, which also notes that efficiency gains could slow that growth. The agentic CPU estimate comes from Arm, “Arm expands compute platform to silicon products in historic company first,” March 24, 2026. The 1:4 ratio comes from Lip-Bu Tan, chief executive officer of Intel, in April 2026. The 8 to 10x figure comes from Rene Haas, chief executive officer of Arm, in June 2026.
An AI agent does more than run a model. Between inference steps it calls tools and APIs, retrieves documents, executes code, and parses results. Those steps run largely on CPUs, one after another. As agents chain more steps together, CPU latency adds up in every response. When an agent browses, fills out forms, or compares prices on your behalf, the latency of each step adds to the time you wait.
“The most important thing is single-threaded performance... If the CPU gets throttled, then we're holding back a whole bunch of GPUs.”Jensen Huang, chief executive officer of NVIDIA, GTC 2026, March 2026
When you open an application, compile code, render a frame, or ask a local AI assistant a question, the speed of one thread sets how long you wait. The same holds on a phone, a tablet, a laptop, or a workstation. Time spent waiting on a computer is time you lose. This kind of compute has improved 4 to 20x over two decades, while parallel workloads gained roughly 4,000x.
Red Hat released a developer preview of Red Hat Enterprise Linux 10 for RISC-V, built with SiFive.
NVIDIA announced that CUDA will support RISC-V host processors.
Ubuntu 25.10 made the RVA23 profile its minimum RISC-V baseline.
Martin Dixon said Google's data centers include x86, Arm, and RISC-V components and set out what datacenter-scale RISC-V requires.
Ubuntu 26.04 LTS carries RVA23 as its long-term supported RISC-V baseline.
The edge systems we target, such as edge servers, networking and storage systems, and high-end embedded platforms, run the same operating systems and software as the data center. That software is arriving on RISC-V. Canonical has made RVA23 the RISC-V baseline for Ubuntu, including its 26.04 long-term support release. Red Hat has released developer previews of Red Hat Enterprise Linux 10 for RISC-V. NVIDIA has said that its goal is to support all major architectures with CUDA, and it has announced RISC-V support.
At RISC-V Summit North America 2025, Google's Martin Dixon said that Google's data centers already include components with x86, Arm, and RISC-V instruction sets. He called for a ratified RISC-V server platform specification and server-class RISC-V chips as the next steps, building on what Google learned moving more than 30,000 applications from x86 to Arm. We see these commitments as the start of a third major architecture alongside x86 and Arm.
Red Hat described its preview in “Red Hat partners with SiFive for RISC-V developer preview for Red Hat Enterprise Linux 10,” May 2025. NVIDIA's statement on CUDA appears in The Register, “Nvidia extends CUDA support to RISC-V,” July 21, 2025. Canonical described its RVA23 baseline in “Canonical and Ubuntu RISC-V: a 2025 retro and looking forward to 2026,” February 27, 2026. Martin Dixon gave the keynote “Paving the Road to Data Center-Scale RISC-V” at RISC-V Summit North America on October 22, 2025, as RISC-V International's speaker listing records. Google reported the 30,000 applications in “Using AI and automation to migrate between instruction sets,” October 21, 2025.
We work with hyperscalers, OEMs, and silicon providers, directly and through ASIC design service partners, on design exploration, evaluation, and licensing.
AheadComputing was founded in 2024 by four CPU architects from Intel. Headquartered in Portland, Oregon, with engineering teams in Austin, Texas and Guadalajara, Mexico, we are working to build the highest-performing, most power-efficient RISC-V CPU cores.
Everybody deserves better compute.
We see a future in which high per-core performance is available to every workload, from hyperscale data centers to the computer on your desk, built on an open standard that anyone can inspect, adopt, and extend. We believe a better computer should serve whoever owns it, whether that owner is a person, an enterprise, or a government. It should put the owner's interests ahead of any third party that profits from the owner's data or dependence, protect the owner's money, time, and attention, and do what the owner tells it to do.
Build the fastest, most power-efficient CPU cores on RISC-V.
We design CPU cores from first principles for data centers and AI, workstations and PCs, phones and tablets, and embedded and edge systems. We intend to show that an open instruction set can lead on performance and on efficiency.
We redesign the core microarchitecture and its goals rather than chase incremental gains.
Every decision answers to the Iron Law, which counts instructions, cycles per instruction, and time per cycle, and to the energy each instruction costs.
We build on the open RISC-V specification. Customers can license our cores to build their own silicon or buy our chiplets, and either way they avoid lock-in to a single instruction set vendor.
Our team has shipped more than 70 products. We hold ourselves to the same standard.
Every product is built on the same RISC-V core. Semiconductor companies can license that core directly or through ASIC design service partners, and OEMs and merchant silicon vendors will be able to buy it as a chiplet.
Our high-performance, power-efficient RISC-V core is available to license for integration into a customer's own system-on-chip or chiplet.
This pre-integrated cluster combines our cores with caches, a coherent interconnect, and debug support, so a customer can drop a complete CPU subsystem into a design.
Our compute chiplet packages the same cores in a system-in-package that OEMs and merchant silicon vendors can buy directly or license.
We plan to extend the same architecture to multi-chip modules with hardware and software designed together, and then to complete systems.
Our founders led CPU architecture at Intel across multiple product generations, from Pentium Pro to Haswell and Skylake.
Debbie and her team at AheadComputing are approaching this problem with the kind of first-principles thinking that actually moves the industry forward… This is the caliber of engineering that delivers generational leaps, not evolutionary steps.Jim Keller, chief executive officer of Tenstorrent and member of the AheadComputing Board of Directors
Since our founding, we have grown to about 120 people, raised $53M, and begun development of our first product.
Four Intel CPU architects launch the company in Portland, Oregon.
The round funds next-generation microprocessor development.
The Tenstorrent CEO joins the Board of Directors.
Eclipse, Toyota Ventures, and Cambium co-lead the round, and total funding reaches $53M.
Dr. Debbie Marr presents “Computing at the Crossroads.”
AheadComputing marks its second anniversary.
You will work with no legacy code and no bureaucracy, guided by the Iron Law of Performance. We are a team of about 120 people building a new CPU from first principles, and every person here shapes the result.
For interviews, executive commentary, speaking engagements, and brand assets, contact our communications team.
We use your email address to send this newsletter and for no other purpose. We do not share it, and every issue includes a link to unsubscribe.
You will work on a clean-slate architecture with no legacy code, alongside a team whose processors at Intel reached billions of users. Join us in Portland, Austin, or Guadalajara.
Tell us about your inquiry and the appropriate team will respond.
AheadComputing develops high-performance, power-efficient CPU cores based on the open RISC-V instruction set architecture. It licenses CPU Core IP and Compute Core Complex IP to semiconductor companies and is developing a compute chiplet in a system-in-package. Its markets are data centers and AI, workstations and PCs, phones and tablets, and embedded and edge systems.
In AI systems, the CPU prepares data, schedules work, and orchestrates accelerators. When the CPU cannot keep pace, accelerators sit idle. Agentic AI workloads also run largely on CPUs. Industry leaders now describe single-threaded performance as a critical factor in AI system throughput.
RISC-V supports both of AheadComputing's businesses. AheadComputing licenses CPU Core IP and Compute Core Complex IP to semiconductor companies and is developing its own compute chiplet. RISC-V is an open specification that anyone may implement without an instruction set license, and AheadComputing's implementation is proprietary commercial technology that conforms to that specification. The software ecosystem's migration from x86 to Arm built the toolchains, memory models, and practices that make a further move to RISC-V easier.
AheadComputing was founded in 2024 by Dr. Debbie Marr, Jonathan Pearce, Dr. Srikanth Srinivasan, and Mark Dechene, CPU architects who led core development at Intel across multiple product generations.
AheadComputing has raised $53M to date, including a $30M Seed2 round announced in January 2026, co-led by Eclipse, Toyota Ventures, and Cambium.
AheadComputing is headquartered in Portland, Oregon, with engineering teams in Austin, Texas and Guadalajara, Mexico.