In artificial intelligence, the most expensive habit is mistaking the current supply chain for a law of nature. Krithik Puthalath has built a company around refusing that mistake. Zyphra, the San Francisco research lab and product company he founded and leads, works on efficient open models, long-memory systems and a cloud platform powered by AMD. Its question is not whether AI will require enormous computation. It is whether every useful route to intelligence must pass through the same architecture, the same hardware and the same gatekeeper.
That question sounds like startup strategy. In Puthalath's case, it also sounds like physics. At the University of Illinois Urbana-Champaign, he studied engineering physics, worked with the Schleife materials-science group and served as an engineering student representative in the campus senate. His name appears on research about silicon nanocomposites and on a later paper that combined first-principles simulations with experiments to study the magneto-optical behavior of antiferromagnetic materials. Before models and GPUs, there were band structures, magnetic symmetries and the quiet discipline of asking what the system is doing underneath the visible effect.
The subject changed. The habit survived. Zyphra's work treats the model, its memory, the training system and the chips below it as parts of one machine. That is a less glamorous pitch than “we have a clever bot,” but it creates more places to be different. Puthalath's public vocabulary returns to efficiency, price-performance and co-design with the insistence of a chorus. The repetition is revealing. He is not decorating the company with an engineering value. He is using that value to choose what the company builds.
Act I - The architecture is a choice
A small model with a long argument
Zyphra surfaced publicly through a series of open research releases. BlackMamba explored a hybrid of attention and state-space modeling. Zamba pushed the idea further: a seven-billion-parameter model with a Mamba backbone and a shared attention block. The architecture aimed to keep the retrieval strengths of attention while reducing the memory burden that grows with long sequences. Its checkpoints were released openly. The technical report said seven researchers trained it over a month for approximately $200,000.
The dollar figure is useful because it makes an abstract belief tactile. Frontier AI discussion tends to use scale as both noun and verdict. A smaller company cannot outspend the largest labs, so it must decide which costs are structural and which are merely inherited. Zamba argued that attention could be used sparingly and repeatedly. One module could do more work. A model could approach peers at its size while requiring less memory during generation. It was not an escape from compute; it was a more opinionated budget.
“Our team is uniquely focused on building the best price-performance multimodal agent systems in the market today.”Krithik Puthalath
There is a founder's lesson hiding inside the architecture. When a market's dominant design makes everyone buy the same scarce input, changing the design can be more valuable than finding a slightly better supplier. It is also harder. Every new choice creates another surface that can fail. Efficient models need data pipelines, kernels, networking and serving software that preserve the advantage. A prototype can win a benchmark and still lose the economics on its way to a customer.
ZAYA1-base: capacity without activating it all
A mixture-of-experts model routes each token through a subset of its capacity. The narrow yellow bar is the point: more stored expertise without paying to activate every parameter every time.
Act II - The chips join the story
An alternative stack, not a consolation prize
By 2025, Zyphra's argument had moved down another floor. Under a multi-year agreement, IBM Cloud prepared a large dedicated cluster of AMD Instinct MI300X accelerators, Pensando networking and ROCm software for the company's foundation-model training. Zyphra then used that integrated platform for ZAYA1-base, an 8.3-billion-parameter mixture-of-experts model with roughly 760 million active parameters.
The milestone mattered beyond one model. Nvidia's advantage is not only silicon; it is a mature ecosystem, familiar tools and accumulated developer confidence. An alternative has to work from compute through networking, then survive the inelegant reality of a large training run. Puthalath called the result a demonstration of co-designing model architectures with silicon and systems. In plainer language, the team did not ask unfamiliar hardware to impersonate the dominant stack. It worked on the seams.
Puthalath's announcement of ZAYA1 also offered a glimpse of how he communicates when the machinery works. He named the research and systems leaders, then the colleagues who drove strategic business development, then thanked the broader research, engineering, product, operations and commercial teams. The public credit was specific. For a project built from dependencies, the thank-you note resembled a dependency map.
“Efficiency has always been a core guiding principle at Zyphra. It shapes how we design model architectures, develop algorithms for training and inference, and choose the hardware.”Krithik Puthalath
His emphasis on partnership appears elsewhere too. Discussing Zyphra's early-stage investor Defined, Puthalath praised how quickly the firm backed the company, but lingered on something more telling: grounded conviction, the right questions and a focus on substance. In infrastructure partnerships, he talks about end-to-end platforms. In financing, he talks about clear questions. The connecting preference is for counterparts willing to understand the whole system rather than admire the visible layer.
Physics leaves the classroom
Puthalath completes engineering physics at Illinois after undergraduate research in materials science.
Zyphra begins
The company forms around efficient multimodal models, novel architectures and open research.
Zamba goes open
A hybrid state-space model makes the efficiency thesis concrete in weights, checkpoints and measurements.
The full stack arrives
A Series A at a reported $1 billion valuation is followed by AMD and IBM infrastructure and the ZAYA1 release.
Research becomes cloud
Zyphra launches an AMD-based platform, beginning with serverless inference for open-weight models.
Act III - Open means operable
The research lab grows a business model
In May 2026, Zyphra brought its infrastructure work to market with Zyphra Cloud. The first service offered serverless inference for large open-weight models on AMD Instinct MI355X hardware supplied through TensorWave. The company also announced fifteen megawatts of MI355X capacity and published a preview checkpoint for ZAYA1-74B, a larger mixture-of-experts model with four billion active parameters.
This is the hinge in Puthalath's story. Research releases attract researchers. A cloud has to satisfy developers and enterprises at the inconvenient hour when a workload is real, the context is long and latency costs money. He described the platform as a natural extension of the research: years spent building, optimizing and validating on AMD, now offered as a service. The statement is almost aggressively practical. The contrarian bet must graduate from proof to utility.
Zyphra's broader mission supplies the reason. The company argues that organizations should control their models, their data and their deployments. Open weights are necessary for that vision, but weights alone are a parcel left at the curb. They need efficient inference, viable hardware choices, deployment tooling and enough capacity to run. Sovereignty becomes credible when a customer can exercise it without assembling a research lab.
The company now describes two mutually reinforcing sides. Zyphra Research works on open multimodal models, long-term memory, continual learning and silicon performance. Zyphra Cloud turns that work into infrastructure for developers, enterprises and AI builders. Discovery feeds delivery; operating the platform reveals the next bottleneck for research. This loop is neat on a diagram and demanding in payroll. It asks one organization to hold scientific taste, systems expertise and commercial reliability at once.
Puthalath's public ambition extends to Maia, a general-purpose agent intended to bring search, communication and productivity tools together for teams. Earlier, he spoke about an operating-system-like layer for devices, controlled through natural language rather than a fixed collection of apps. The product names may evolve, but the aspiration is consistent: models that remember, continue learning and become useful across modalities, while remaining controllable by the people and organizations using them.
“We've spent years building, optimizing, and validating AI systems on AMD infrastructure, and are now bringing that capability to market.”Krithik Puthalath
The physicist's advantage
It would be tidy to claim that studying antiferromagnets inevitably leads to building an AI cloud. Careers do not obey equations so politely. The useful connection is subtler. Physics trains attention on constraints, interactions and scale. Puthalath's company applies those instincts to a market inclined to discuss model capability as if it floats above hardware. At Zyphra, the chip matters. The network matters. The memory cost matters. The license matters. The route from laboratory result to customer workload matters.
That worldview creates a distinct kind of founder story. Puthalath is not a celebrity CEO with an archive of lifestyle anecdotes. His public record is largely the work itself: papers, model reports, partnership announcements and precise acknowledgments of colleagues. The personality available in those artifacts is methodological. Ask the right questions. Focus on the substance. Credit the people who carried the system. Measure efficiency. Keep the architecture negotiable.
Zyphra still faces the severe test attached to every full-stack ambition: each additional layer creates another business capable of consuming the whole company. A research lab, a model family, an agent product and a cloud platform could each be its own demanding enterprise. Their value together depends on integration producing leverage rather than distraction. The elegance of the thesis will not make the operations easy.
Yet the shape of the wager is now clear. Puthalath is building for a future in which advanced AI is open enough to inspect, efficient enough to deploy and portable enough to avoid a single infrastructure path. He has moved from studying how material systems respond when symmetry is disturbed to asking what happens when the AI industry's symmetry is disturbed too. The answer will arrive in benchmarks, invoices and whether developers choose the other door.