2025: The Year We Moved from Promises to Proof
If 2024 was about building, 2025 was about proving.
This year, NextSilicon crossed a line that every deep-tech company dreams about: we moved from “interesting ideas” to “real systems, real workloads, real results.” In October, we publicly disclosed Maverick-2’s architecture and performance, including out-of-the-box acceleration and industry benchmark highlights, plus the broader story behind our Intelligent Compute Architecture (ICA)TM.
And the timing could not be more telling.
When NVIDIA makes a huge bet like the reported ~$20B Groq acquisition, it's an acknowledgement that GPU architectures aren't one-size-fits-all. Despite their training dominance, as AI workloads diversify rapidly across training, inference, and reasoning regimes, they will demand different optimization points. Each architecture has its sweet spot, so the industry stitches together the right compute for the right phase via software and interconnect. But here's the real question: what if reconfigurability was built in from the start?
I remember a conversation with an investor during our Series A back in 2019. He told me, essentially: “Why do we need a general-purpose accelerator? It’s all matrix multiplications.”
Fast forward a few years, and reality is more interesting.
Yes, modern AI needs matrix multiplication, but it also needs pointwise kernels, reductions, memory movement, and interconnects. And if you care about real efficiency, you end up caring about the messy parts, not just the headline GEMMs.
What 2025 clarified for me
For our first 6.5 years, we stayed unusually quiet, deliberately.
Until you deliver silicon, most of what you say is storytelling and hand-waving. Sometimes it’s necessary, but it’s still not the thing. The thing is running code.
1) “Silicon, not slideware” is not a slogan. It’s a survival skill.
We are shipping real hardware and publishing real data, not telling a story that only exists in a deck. We shared benchmark highlights, talked about unmodified code paths, and anchored it in production reality, including deployment at Sandia.
A founder learns fast: hype has a short half-life. Measurement lasts.
2) Developer time is the most expensive compute on Earth.
The ugly truth of accelerated computing: huge fractions of HPC effort are burned on porting, rewriting, and tuning instead of science. I’m still obsessed with this. If “acceleration” requires months of re-engineering, it is not acceleration. It’s a tax.
3) HPC and AI are converging, but HPC still keeps us honest.
AI workloads, in a strange way, are simpler compared to the vast mathematical diversity of HPC. AI has a smaller set of dominant primitives that repeat at a massive scale. HPC is different. It’s every kind of kernel, every flavor of sparsity, every kind of irregularity you can imagine. All of it demanding correctness that can’t drift over millions of timestamps.
That’s why HPC is the best lie detector.
FP64 is still the backbone of real science and always will be. Precision is not nostalgia. It is how simulations stay stable and how reality doesn’t drift into numerical chaos. You can't hand-wave your way through double precision. You can't fake broad kernel support. You either run the code correctly or you don't.
My take: the future belongs to architectures that can handle the throughput demands of AI and the correctness demands of HPC on the same platform. That's not a nice-to-have. It's table stakes.
Dataflow, in my opinion, is the best architecture for doing that. Not because it’s trendy, but because it has the flexibility to run different workloads efficiently without drowning in control overhead. To run matrix-heavy code and the weird, messy kernels that nobody wants to talk about but that scientists actually need. If you want one architecture that works across both worlds, maximum flexibility isn't a feature. It's the foundation.
The year, as a story (and a little nerdiness)
At the start of 2025, the theme was already clear: compute is becoming the bottleneck for discovery. We are generating data faster than we can turn it into insight. Every lab, every simulation, every instrument is screaming the same thing: “I can produce more than you can process.” And the painful part is that the answer is rarely “add more hardware”. The answer is “make it usable”.
That’s what made me bring up Bring Your Own Code (BYOC) multiple times throughout the year. Bring your own code should be the default expectation, because science should not have to rewrite itself to fit hardware. The entire premise is backwards when researchers have to become full-time porting engineers.
Somewhere mid-year, I found myself thinking about something more personal: why I care so much about usability and about the “it just runs” experience. The answer is simple and slightly embarrassing: I was formed by debugging. By long nights of chasing bugs that felt irrational until they suddenly made sense. You don’t forget the feeling of a system punishing you for trying to use it. And you definitely don’t forget the moment when you finally tame it.
Then came the public moments. Paris, RAISE Summit, conversations where AI’s appetite collided head-on with Moore’s Law fatigue. The kind of events where you can witness an industry shifting in real time. Even the jokes, like showing up with Commodore T-shirts, carried a point: compute has always pivoted when the old assumptions stopped working.
And then the year’s turning point: Maverick-2 went public.
That shift from building quietly to showing the world is not a marketing moment. It’s a psychological one. For years, you know what you’re building. Your team knows. A handful of partners know. But the broader world only knows what it can run, test, and measure.
Once you cross that line, everything gets sharper. The questions become better. The feedback becomes more useful. And the only currency that matters becomes real results.
SC25 was the icing on that story. Recognition is nice, but what mattered more was what it represented: the community that has seen every hype cycle is paying attention to substance again. That’s the only kind of attention worth earning.
What I’m Taking Into 2026
Make adoption boring. The best compliment is “it just ran.” BYOC is not a feature. It’s the point. Success means infrastructure that disappears into the background so researchers can focus on science, not our systems.
Keep the bar anchored in science. FP64 correctness, real workloads, and honest benchmarking will keep guiding us. Enabling scientific discipline remains our north star, HPC won't let you fake it.
Build for change. Workloads will evolve. Models will change. “Future-proof” is not marketing, it’s an engineering requirement. The architectures that last are built to adapt.
Scale the humans. The team is the product. Belgrade is one example of how seriously we’re investing in that. Scaling the organization with the same intention as the technology sustains long-term impact.
Gratitude
To my co-founders Eyal Nagar and Ilan Tayari, and to the entire NextSilicon team: you earned every bit of this year.
To customers and partners who put real workloads on real silicon, and to the HPC community that rewards substance and ignores noise: thank you.
Onward.
Stay Connected
Stay tuned as we share more details in future blogs and announcements via our Nextletter.
About the Author:
Elad Raz is the founder and CEO of NextSilicon, a company pioneering a radically new approach to HPC architecture that drives the industry forward by solving its biggest, most fundamental problems.
