In 2004, we made an early bet on the National Stock Exchange of India. The conviction was simple: India deserved electronic trading rails, not a trading floor. NSE's daily cash turnover has since grown nearly 18x. Better rails brought in more investors, and those investors got more sophisticated as the market grew around them.

‍

We've seen that pattern play out in every market we've watched. Lower the friction, and more people show up. Give them better tools, and the market gets smarter with them.

‍

So here's a question. You think a super El Niño is coming this winter. What's the trade? Coffee? Cocoa? Fertilizer stocks? A prediction market on hurricane counts? Which venue? What size? What's the hedge? When do you get out? By the time you've worked through all of it, the move already happened. Most people never place the trade at all.

‍

That's what Catalyst fixes.

Why is the trade still handmade?

Every generation of trading killed exactly one friction. Screens killed the floor. E*Trade killed the phone call to your broker. Robinhood killed the commission. Newer venues killed market hours and opened up asset classes that didn't exist ten years ago.

‍

Nobody killed the hardest one: turning an idea into a position. Ideas are fuzzy, qualitative and cross-asset. Execution is mechanical and venue-specific. At a hedge fund, that gap gets closed by analysts, quants and an execution desk. Everyone else closes it with a spreadsheet, five browser tabs and hope.

‍

We see this gap up close. As a crossover firm, Premji Invest runs a public markets team alongside our private investing practice. Every day, that team does the work Catalyst is automating: turning a view on the world into positions, sizing them, and watching for the signal that says the thesis has changed. That gave us an unusual lens in diligence. We weren't just evaluating Catalyst as software. We were asking whether it could do the job our own investors do, and where it would break.

What is Catalyst?

Catalyst is the execution desk most people never had. You type what you believe, literally "El Niño, what's the trade?", and Catalyst's agents research the idea, pick the instruments and venues that best express it, and hand you a strategy you can read before anything runs.

‍

And it doesn't stop at the order. Each strategy is a standing workflow that watches the market for you. If a major outlet reports X, take the position. If another contradicts it, unwind. Take profit here, stop-loss there. Prediction markets, equities, options, perps, sentiment, fund flows and news, all in one interface.

‍

Why not just let the AI trade?

‍

This is the part of the company we spent the most time on in diligence.

‍

Most AI trading products we've seen put the model in the loop. Every time the market moves, the LLM looks at the state of the world and decides what to do next. That's fine in a demo and a problem with real money. Inference on every tick gets expensive fast. Models still fumble a small share of tool calls, and on a $100K account even a rare mistake is a lawsuit. And when the agent sells something at 3am, there's no clean way to reconstruct why.

‍

The fallback most teams reach for is a library of canned strategies. Safe, but nobody's actual thesis fits in a template.

‍

Catalyst does something different. The model writes the strategy, and then it's done. When you create a strategy, the LLM turns your idea into a program, a fixed set of signals, thresholds, venues and actions, that you can read before it ever touches your money. From then on, that program runs on its own. Dylan's framing is that models are great at writing code, so you should treat every agent like a script.

‍

That choice drives the economics too. Catalyst pays for intelligence once, when a strategy is written, not every time it trades. Cost grows with the number of ideas on the platform, not with volume. Today, the inference bill for the whole platform rounds to zero.

Whose edge is it?

There's a second problem with most AI trading tools, and it's a bigger one. LLMs are built to give the most likely answer and to agree with whoever is asking. That's the opposite of what a trader needs. If a million people ask a chatbot for the El Niño trade and get the same answer, that trade is crowded before anyone places it. Consensus isn't alpha.

‍

Catalyst starts from the other end. The thesis is yours. Catalyst's job is to sharpen it and carry it out, not to swap it for the internet's average opinion. It puts the data in front of you, from prediction market odds to positioning, options flow and sentiment, so you can see what the market already believes and whether your view is actually differentiated. Then it executes what you decided, not what the model would have preferred.

‍

The best traders we know don't want an oracle. They want their own judgment, executed faster and more precisely than they could manage by hand.

What's the catch?

If this were easy, every brokerage would have shipped it already. Three things make it hard.

First, data. There are maybe twenty core actions in all of trading: long, short, swap, hedge, rebalance. Strategies don't win on exotic execution. They win on signal: order-book depth, an obscure news thread, crop yields, satellite imagery. Catalyst pays for that long tail and wraps it in its own query language so agents can compose across all of it. Every strategy users build teaches Catalyst which signals actually matter. That's the moat, and it compounds.

‍

Second, trust. Nobody hands their money to a black box. Catalyst never takes custody. Keys sit inside hardware-isolated enclaves that Catalyst's own code can't read, and every trade executes at an established third-party venue or broker. Catalyst is software, not a broker.

‍

Third, incumbents. Big brokerages could build agentic features. But a neutral layer that routes across dozens of venues and asset classes is genuinely hard for any single platform to build. In Catalyst's architecture, every venue is just a swappable adapter. Our bet is that a lot of incumbents end up as distribution partners, not competitors.

Our thesis

The consensus take is that AI in trading is a feature. Incumbents bolt a chatbot onto the app and call it a day. We think that's wrong.

‍

Interface shifts crown new winners. The floor brokers didn't build E*Trade. E*Trade didn't build Robinhood. Every time the interface changed, the company that defined the new one won the next generation of traders. The agentic interface will be no different, and it'll be won by whoever makes AI trustworthy enough to touch your money.

‍

The twist is that Catalyst doesn't have to beat the venues to win. It routes to them. Twenty years ago, we backed the rails. Now we're backing the interface that sits on top of all of them.

The team, and the signal

This is one of the spikiest teams we've met. Dylan Iskandar was an AMP Scholar at Jane Street, built and ran his own trading operation before he finished high school, and did formal verification research at Stanford before dropping out to start Catalyst. He's lived the exact workflow he's automating, and his obsession with provable, deterministic systems shows up everywhere in the product. JZ (Justin Zheng) worked on the early protocol that became Worldcoin as a teenager and is one of the most connected operators of his generation in the Valley. The two met through Tyler Cowen's Emergent Ventures when Dylan was in tenth grade. The rest of the team is Math Olympiad and quant-firm alumni.

In Closing

Every shift in trading removed one layer of friction. Catalyst is going after the last and hardest one: building the trade itself. We're thrilled to back Dylan, JZ and the team alongside Sequoia, Jump Trading, Selini Capital and a bench of traders and AI researchers who are also some of Catalyst's earliest users. Onwards.

Premji Invest is excited to announce our investment in CScale, which is building a new optical interconnect architecture for scale-up networks.

‍

We have been consistently impressed by the depth and breadth of CScale's team, particularly their focus on the hardest problems gating broader adoption of optical interconnects. Martin Lund brings decades of experience building networking silicon and infrastructure at datacenter scale, while Sanjai Kohli has spent his career building communications and computing systems. The broader team spans high-speed electronics, photonics, packaging, and software. That range of experience gives the team the perspective to approach the optical interconnect problem from first principles and design around what it actually takes to deploy these systems commercially.

‍

As datacenters scale, interconnect architects must design for higher throughput and bandwidth density within tighter reliability, power and cost constraints. While scale-out networks have already moved substantially towards optics, scale-up fabrics remain largely electrical. However, as copper reaches its physical limits, optical scale-up will become increasingly necessary to sustain further power and cost improvements in future rack architectures. Unlike copper, fiber offers a vast pool of spectrum allowing more throughput within the same physical footprint.

‍

Our work across the category has led us to a broader conclusion that commercial viability will not be defined by a single component-level innovation alone. This is fundamentally a systems problem, where manufacturability, serviceability, reliability, and system-level operation need to be treated as primary design requirements alongside raw performance.

‍

The industry has historically been undercapitalized, but it has made meaningful progress over the past few years in advancing the capabilities of optical devices. As such, we expect large customers to experiment with a range of approaches, both internally and through external partnerships, before committing meaningful commercial volumes to any one solution. With that in mind, we evaluated the category through a few first-principles checks.

Components must be co-optimized and integrated as a complete system to support the requirements of mass deployment

Several existing approaches in the industry focus on individual components (e.g. faster modulators, higher-powered lasers, or differentiated material systems). But as we worked through the requirements of deployment at scale, we found that preserving component-level advantages can require burdensome tradeoffs elsewhere in the system. For example, a highly performant modulator may compromise packaging yield or thermal stability. A novel, power-efficient light source may introduce crosstalk when densely integrated. A novel material platform may prove difficult to yield at high volumes. Conversely, a disaggregated approach shifts the burden of integration, validation, serviceability, and support to the operator, adding substantial operational overhead and technical complexity.

‍

One architectural path we examined relies on dense arrays of optical resonators such as microring modulators. In isolation, these resonators look attractive because of their compactness and efficiency. The challenge is maintaining stability as these resonators are densely integrated. In this format, thermal crosstalk and wavelength drift become a coupled, system-level control problem, increasing the demands on tuning, calibration, and thermal management. In our view, this may not be the optimal tradeoff for deployment at scale. The key question is whether these architectures can be manufactured at high yields and perform reliably in the field, especially as designs become densely multiplexed and production volumes rise.

‍

The optics industry has learned many of these lessons in past cycles. Technologies can appear compelling in an early demo or at the subsystem level, but become harder to qualify, manufacture, and operate reliably in the field. Our work reinforced that in photonics: strong component-level performance may not be enough to drive commercial adoption.

‍

We saw much of the same thinking in our early conversations with CScale. The team's approach began with the full set of requirements for a scale-up fabric that could actually be deployed and operated at scale. The result is a product that combines proven photonics process know-how with an architecture designed around the practical realities of optical systems, including thermal instability, aging, process variation, and component failure. Manufacturability and serviceability are built into the design from the start. That approach closely matched the conclusions we had reached through our own work and further underscored CScale's differentiation.

Reliability is mission-critical for scaled commercial deployments

Bandwidth, power, and reach tend to dominate discussions around new optical devices. We think reliability is just as important, and one of the biggest challenges to solve for broad adoption of scale-up optics.

‍

At datacenter scale, optical components are inherently difficult to manufacture and operate reliably. Unlike CMOS, optical systems require different materials and components (e.g. lasers, modulators, fibers) to be integrated and precisely aligned across multiple manufacturing steps, many of which can be difficult or impossible to rework. Small alignment errors or wavelength shifts can degrade link performance, while adding more components can compound yield and reliability challenges.

‍

Those challenges do not end at manufacturing. Optical component behavior is sensitive to environmental conditions, and failures can be difficult to isolate or reproduce. As optics moves closer to the compute package, engineers cannot depend on flawless operation of every component. The system must be able to monitor link health, diagnose degradation, and continue operating when individual components fail.

‍

The cost of failure is especially onerous at cluster scale. Low failure rates at the component level can create a substantial operational burden across the enormous number of links in a modern AI cluster. And because these workloads are tightly coupled, a single degraded link can create cascading stalls and significant losses in compute utilization. A clean eye diagram in the lab or a strong component-level reliability result is not enough for commercial viability. What ultimately matters is whether the full system can be relied upon under real operating conditions.

‍

CScale has designed its system with these challenges in mind across the full product lifecycle, from manufacturing and initial deployment through real-time optimization and failure management. The goal is higher cluster availability, simpler operations, and ultimately more confidence for engineers to scale beyond the limits of legacy interconnect.

Spectrum is abundant

At a high level, the throughput of an optical link depends on two things: how much spectrum is used and how many bits are transmitted through each unit of spectrum. The industry has made strong progress in advancing spectral efficiency, and in datacenter optics, these gains can help address the density constraints imposed by limited package perimeter and shoreline availability. However, we believe that overly relying on higher symbol rates and more complex modulation may be the wrong approach for short-reach scale-up. Accordingly, we see potential in product roadmaps which can more effectively exploit the vast spectral headroom provided by optical fiber within the packaging and power constraints of the system.

‍

In a medium with abundant spectrum, maximizing bits per second per hertz should not necessarily be the primary goal. Denser modulation or higher symbol rate scan make signals more fragile, consume more power, demand heavier error correction, and reduce tolerance for thermal drift, crosstalk, manufacturing variation, or loss. In short, higher-throughput lanes do not necessarily result in a more robust system. One logical alternative is to use more of the available optical spectrum while keeping each channel simpler and easier to manage. The opportunity is to trade abundant spectrum for better signal integrity, lower latency, and greater operating margin.

‍

Ultimately, CScale's technical depth and systems design judgment were central to our conviction. As optics moves into the scale-up fabric, we believe commercial viability will depend increasingly on building reliable, well-optimized systems, not just higher-performing components. We are excited to partner with CScale as the team builds for that future.

When we made an early investment in the National Stock Exchange (NSE) back in 2004, we had a clear conviction - to drive the country's growth potential, India deserved a powerful electronic equity trading architecture with a multitude of quality instruments.

‍

Between 2000-2005, NSE had launched 1) index options based on the NIFTY50 index for trading, 2) single stock futures and options on listed securities, 3) ETF listings, and 4) NIFTY bank Index derivatives (*). NSE's pioneering actions were clear signs of an emerging leader that supported our conviction.

‍

To give some perspective of the journey we saw: the average cash daily turnover of NSE has grown to INR 1130 billion (2024-25) from just INR 63 billion (2005-2006); the number of securities grew to 6,381 from 1,288 in the same period(**). Market maturity, proactive regulation and policy changes supported; as did retail participation and digital adoption, causing transaction volumes to explode. This journey was certainly not without its ups and downs with global and local events pulling against India's potential.

‍

Our evergreen, patient capital allowed us to stand firm alongside the business without wavering, proving that long-term backing yields disproportionate success.

‍

Today, NSE’s pre-eminent standing is validated by an impressive, world-class roster of anchor and institutional investors. Marquee global and domestic institutions—including the Life Insurance Corporation of India (LIC), Monetary Authority of Singapore (MAS), Abu Dhabi Investment Authority (ADIA) and Goldman Sachs serve as a testament to the exchange's incredible success and systemic importance. NSE has seen overwhelming investor interest for debut on bourses with a market cap of approximately INR 4.4 trillion at the upper end of price band and over subscription by 5.7x, making it amongst the largest demand generators in the recent times for OFS on listing.

‍

Partnering through this monumental journey has provided us with invaluable structural insights into India's financial services development and a blueprint for navigating complex regulatory landscapes. We are actively channeling this deep operational fluency into our next frontier: backing early-stage deep tech business models, tokenization, and next-generation financial rails.

‍

For visionary founders building in heavily regulated, technically complex spaces, our track record shows we don’t just fund such vision—we stand beside you and actively support your journey to reach the inflection point.

‍

(*) History & Milestones - NSE India;

(**) Live Analysis - Capital Turnover & Business Growth - NSE India

‍

‍

Few companies in AI-driven biology have carried a molecule of their own into a patient. Fewer still have done it on a format most of the field has not attempted. BigHat Biosciences has done both, and today it announced a $75 million Series C, co-led by Premji Invest and DFJ Growth, to do it again. 

‍

BigHat is an AI-native biotech that pairs frontier AI with an autonomous, high-throughput wet lab to design and develop next-generation protein therapeutics. Its platform has now produced two wholly owned programs: BHB810, a CDH17-directed antibody-drug conjugate for gastric and other gastrointestinal cancers that is already in patients, and BHB299, an avidity-driven CEACAM6 T-cell engager for solid tumors that is completing preclinical development, ahead of human trials in early 2027. Around them sit collaborations with Amgen, Merck, Johnson & Johnson, AbbVie and Eli Lilly, and partnerships with leading frontier AI companies. Biologics are roughly half of the addressable drug market, and BigHat is one of very few companies that can point to a clinical pipeline, validation with pharma and frontier AI companies, and a proprietary integrated data engine at the same time.

Why biologics fail, and why that is the opportunity

Biologics rarely fail in the clinic because they cannot bind a target. They fail because they lack the properties that turn a promising binder into a safe and effective medicine: internalization, selectivity between tumor and healthy tissue, a safety window wide enough to dose, and the developability and manufacturability that let a molecule survive CMC and an IND. The industry has spent tens of billions of dollars learning this. Rova-T died on payload toxicity, TGN1412 on cytokine release, solitomab on a therapeutic window that never existed. In each case the target was right and the molecule did what it was designed to do. What failed was a property nobody had written into the specification.

‍

BigHat treats binding, developability and function as one joint design challenge rather than three sequential ones, so that the biological need and the modality built to meet it are aligned from the first design cycle. The properties that kill drugs are written into the objective from day one, engineered in on every molecule at scale, and measured on purified protein rather than inferred. Crucially, that matters more now than it did a decade ago. Most of the targets reachable with a conventional monoclonal antibody have been picked, and the next phase of drug discovery depends on formats and functions that antibodies have not traditionally had: conditional activation, avidity-driven selectivity, payload delivery, logic gating across two antigens. Those are formats where a good binder is the starting point rather than the answer, and BigHat is purposefully built to design for them. That is the discipline of a drug company applied at the speed of a machine learning loop, and it is why BigHat has a molecule in patients. 

How Premji Invest evaluates AI-native drug discovery companies 

At Premji Invest, we spend a lot of time on AI-native drug discovery, and our view has narrowed to something fairly specific. General-purpose models and agents are commoditizing quickly; a state-of-the-art model gives you a head start, not a business. What does not commoditize is the closed loop of design, make, test and learn, the proprietary data it generates on hard problems, and the clinical judgment that decides which problems are worth solving. When we underwrite a company in this space, we focus on four things: 

• Loop velocity. How fast does a design become data, and how fast does that data retrain the model? A platform’s real value is its learning rate. 
• Proprietary data on hard problems. Is the company generating fit-for-purpose, multi-parameter data on molecules that are trying to become drugs, or fine-tuning on binding data everyone else can download? 
• Decision advantage. Can it kill inferior candidates in silico and in weeks, before spending months and millions in the lab and years in the clinic? 
• Dated clinical milestones with a capital path to reach them. In biotech, undercapitalized programs fail about as often as bad science does. 

BigHat is the clearest expression of that framework we have found in biologics. 

The loop: One week from design to data

What slows AI down in biology is the absence of task-specific, high-quality data. BigHat solves this by making its own, at scale. 

1. Integrated AI and wet lab. BigHat’s platform processes 2,000+ antibodies per week and measures more than ten properties on every one of them, on the same purified molecule. Paired multi-parameter measurements are a different and far more useful training object than a pooled binding screen, because a frontier model designing a drug has to respect several design constraints at once. 
2. Weekly cycles. BigHat goes from design to data in about a week, an industry-leading cycle time that supports roughly 40 design-make-test-learn cycles a year against the two to four a traditional group manages. Every cycle reduces model uncertainty on exactly the question that matters. 
3. Failure data. Nobody publishes the molecule that aggregated, or the format that lost function on reformatting. BigHat runs those molecules, records why they failed, and feeds the answer back into the next design cycle. Public and partner datasets are almost entirely positive results; BigHat’s includes the negatives, which is where most of the learning is.

This is the piece we keep coming back to. No company measures self-association, viscosity, CMC-scale expression or tox-relevant behavior unless it is trying to make a drug. This dataset is downstream of BigHat’s clinical discipline rather than a substitute for it, and that is why BigHat’s dataset advantage compounds with every program the company runs. 

What the loop makes possible

Speed is only half of it. Running the loop this fast changes what can be designed at all, because it lets you optimize for many properties at once instead of trading them off in sequence, and it lets you take on formats where design quality determines whether a target is druggable at all. 

• Hard problems, not repeated problems. BigHat’s programs each solve a different class of design constraint: an ADC that must be internalized and trafficked correctly to release its payload, a T-cell engager whose selectivity comes from avidity and two-arm geometry rather than from either binder alone, and logic-gated multispecifics that fire only in the presence of the right combination and stoichiometry of antigens. Binder generation has advanced enormously across the field – BigHat goes beyond it.
• Multi-objective optimization. Affinity, stability, expression, aggregation, immunogenicity and function are optimized jointly, which is how you get a developable drug that works as intended.
• Complex modalities. Multispecifics, T-cell engagers and ADCs: the formats where sequence-only models have the least to say and where almost no public training data exists. 

Pharma partnerships and a wholly owned pipeline 

BigHat pairs a wholly owned pipeline with pharma partnerships, which deliberately blunts the binary risk of an asset-centric biotech and provides the most credible external validation available. 

• Partner of choice. Collaborations with Amgen, Merck, Johnson & Johnson, AbbVie and Eli Lilly, including a completed strategic collaboration with J&J, three project collaborations with Merck, and an expanded relationship with Lilly that now includes machine learning-enabled biologics discovery through Lilly TuneLab. In a market where AI discovery deals still command only a small share of total deal value upfront, repeat business and expanded scope from partners of this caliber is the signal that carries weight.
• Frontier AI partnerships. BigHat has also formed partnerships with leading frontier AI companies to generate high fidelity, fit-for-purpose training data and deploy state-of-the-art models for agentic therapeutic design. The reason those labs come to BigHat is instructive: binding is close to a solved benchmark, and the unsaturated question is whether a model can design molecules that survive everything else. BigHat is one of the few places where that question can be scored at scale. 
• A pipeline built on validated biology. BigHat applies the platform to targets where the biology is well understood and earlier programs failed on molecular properties rather than on mechanism, so the variable being tested in the clinic is design quality. BHB810 dosed its first patient in July 2026; BHB299 is expected to enter the clinic in 2027. 

‍

That last point is why the timing matters to us. The market still prices AI-native pipelines at legacy biotech success rates, even as early evidence suggests AI-discovered assets have cleared Phase 1 at higher rates than the historical average. We think that gap closes as clinical data accumulates, and we would rather own the company generating that data than wait for the consensus to move. 

What’s next

Under CEO Peyton Greenside, BigHat has assembled a team that is rare in this field: protein engineering, AI, lab automation and clinical development under one roof, with the experience to carry molecules through GMP, an IND and a first-in-human study rather than hand them off. Becoming a clinical-stage company is a step change, and it puts BigHat at the front of the field at exactly the moment the field is being asked to prove that AI-designed biologics can become successful medicines. 

‍

At Premji Invest we seek to partner with teams that think big and have the discipline to build enduring companies. BigHat is exactly that, and we are proud to co-lead this financing alongside DFJ Growth and to support Peyton and the team as they bring transformative medicines to patients. 

‍

If you are building a company at the intersection of AI and therapeutics, we would like to hear from you. 

‍

We are thrilled to announce our investment in Databricks and to partner with Ali Ghodsi and his entire team. Databricks sits at the center of the most critical architectural shift since the cloud: the transition from human-operated to agent-operated software. We believe they have built the only platform designed from first principles for this era.

‍

At Premji Invest, our approach is rooted in developing a deep understanding of the product vision. When a technology stack shifts, we dig deep to understand where bottlenecks may develop and what capabilities can accrue value at scale. As enterprise agents move from concept to production, our thesis is clear: Databricks is evolving from a leading data and analytics platform into the foundational execution and governed context layer for the agentic enterprise. Critically, this is not a pivot away from the core. We view the lakehouse business as one of the largest beneficiaries of this shift across software, and it is reaccelerating rapidly as a result. On top of that base sit three strategic growth levers that together fulfill the vision of becoming the data fabric for the enterprise in the AI era: Lakebase for transactional workloads, Genie and the agent layer that governs and routes agentic work, and a new class of agentic applications built directly on the platform.‍

Enterprise AI has a context problem‍

Foundation models have become remarkably capable: they can write code, reason across multiple steps, and interact with external tools. However, intelligence alone cannot automate real enterprise work. Many AI agents fail in production due to poorly engineered context and missing access controls.

‍

Historically, the enterprise data stack was built for human behavior: analytical query volumes scale with the number of data analysts, administrators pre-provision databases, and compute cycles are tied to rhythms of the human workday. Agents fundamentally break this paradigm. They issue exponentially more queries, spin up databases on demand, and burst compute 24/7.

‍

Furthermore, agents lack the implicit context human employees possess. For example, they do not intuitively know which of the seven revenue definitions is correct or what the business means by "active customer." Letting agents reason over raw tables at runtime is slow, expensive, and unreliable. The true bottleneck is governed context: the accumulated semantic knowledge of the organization's entire data estate.

‍

Agents need three things that the historical enterprise data stack does not provide:

1. An operational foundation built for programmatic, bursty workloads
2. Accurate answers grounded in governed context
3. Economic control, i.e., forecastable AI budgets and the ability to route each task to the most cost-effective models (vs. sending all tokens to expensive frontier-models)‍

‍Unity AI Gateway and Genie Ontology‍

Databricks addresses this challenge through Unity AI Gateway, which helps companies control AI access, spend and observability across agents, tools, models, and MCPs. On top of it, Databricks has built the Genie Ontology, a live context layer and continuously constructed knowledge graph that extracts metric definitions, business terms, and asset relationships from the customer's own tables, queries, and documents.

‍

This context is precomputed rather than retrieved at inference time. Genie Ontology does the extraction and ranking work continuously in the background, so at runtime, the agent reads a summary of what the organization knows instead of reasoning over raw data (expensive and slow). This approach has proven to yield significant cost and latency gains while simultaneously providing the fine-grained controls to securely make the enterprise data accessible to AI.‍

Databricks will monetize the entire agentic data stack across transactional and analytical workloads‍

As agents collapse the cost of creating software, the number of applications inside an enterprise is increasing exponentially. With agents, these applications are being instantiated on demand with unpredictable access patterns. Databricks' Lakebase is a modern take on OLTP that we believe is best positioned to be the default operational backend for this wave.

‍

Lakebase is built on Postgres, the open-source database that has become the default foundation for modern applications and, increasingly, for AI agents. Postgres stands out for its customizability, extensive ecosystem of drivers, extensions, and migration tooling, along with the operational knowledge already sitting inside every engineering organization. Lakebase separates compute from storage, moving durability into object storage so the compute layer holds no durable state of its own. That is what lets Lakebase scale compute to zero when idle and come back on demand, which is exactly what agent workloads require. Agents can create, use, and abandon databases dynamically without anyone paying for provisioned capacity in between, and can safely test changes against instant, disposable copies of production data. Roughly 80% of databases on the Lakebase architecture are already created by agents rather than humans.

‍

We see a growing number of workloads migrating to Lakebase because it is the only operational database that agents can provision programmatically inside a platform the enterprise has already governed. Beneath the operational data sits a larger body of information: documents, events, logs, traces, and embeddings. Because this operational context lands in an open format under the same catalog and storage as the rest of the company's data, it is immediately available for analytics and AI in the Lakehouse without having to build and manage brittle pipelines.‍

Growth durability: Databricks is an index bet on enterprise AI‍

We believe we are in the extreme early innings of enterprise AI adoption. Production deployment is constrained by data readiness, permissions, workflow redesign, reliability, and org change. Goldman predicts that enterprise agents alone (excluding consumer usage) could account for ~12x increase in token consumption vs. all tokens consumed today across enterprise and consumer.

‍

Conventional forecasts underestimate Databricks by extrapolating solely from today’s human-driven workloads. We believe with high conviction that the market is getting this wrong, and that Databricks is positioned for faster, more durable growth.

‍

Databricks does not need to predict which foundation model, agent framework, or end-user interface ultimately wins. As agentic activity scales, consumption cascades down the stack: agents require data ingestion and transformation (data engineering), data preparation and low-latency compute across heterogeneous data types (Lakehouse//RT, Photon, Spark, Delta, Iceberg), lifecycle governance (MLflow), orchestration (Agent Bricks), and post-execution analytics and reporting (SQL). Databricks monetizes the compounding compute cycles across the entire data estate.

‍

Thus, every successful agent creates more demand for the core  governance stack (Lakehouse, Data Engineering, SQL, ML) while accelerating adoption across all three growth levers: Lakebase in transactional workloads, the Genie and agent layer that routes and governs that work, and the agentic applications Databricks is now building directly on its own platform, beginning with customer data and security.‍

Partnering with a generational team‍

Ultimately, our conviction in Databricks rests on the team. Very few technology companies successfully crest more than one secular wave. Ali Ghodsi and the Databricks leadership have compounded execution for over a decade while repeatedly expanding the company's footprint into new categories: from big data with Spark, to the lakehouse architecture and data warehousing, to GenAI, and now to OLTP and the agent layer with model serving and routing. Databricks has won in each on the strength of a single coherent platform rather than a portfolio of disconnected products. What stands out most is that these waves are being caught nearly simultaneously rather than in sequence, which is what separates Databricks as an enduring franchise. We believe Databricks will be the data and governance foundation on which the agentic enterprise is built, and we are honored to partner with the Databricks team on that journey.

‍

We're excited to lead Upscale AI's Series A1 round and accelerate the company's next phase of growth. This round deepens our partnership with a team we've long believed in. Upscale AI's founders previously architected Innovium, and we backed them then, ahead of its $1.1B acquisition by Marvell. Building leading-edge networking silicon is one of the hardest and least forgiving disciplines in semiconductors, and this team has already done it at scale, in production. Just as importantly, Upscale AI brings together world-class silicon, systems, and software talent, with proven experience from Cisco, Juniper, Palo Alto Networks, and other leading infrastructure companies.

‍

Upscale AI's collaboration with NVIDIA cements its role as a key networking partner for next-generation AI data centers, enabling heterogeneous compute at scale. By leveraging standards-based Ethernet, AI builders can seamlessly integrate custom XPUs into high-performance AI clusters without sacrificing network efficiency.

‍

Upscale AI's focus on Ethernet and open standards aligns with the market's shift toward open architectures. We are also compelled by Upscale AI's software strategy. Rather than weighing networks down with legacy enterprise features, Upscale AI is building a purpose-built network OS optimized for AI clusters, improving performance and reducing complexity. This creates a tighter link between the needs of AI infrastructure teams and the capabilities of the network.

‍

The AI infrastructure market is moving quickly, and companies that can work across major silicon, cloud, and systems partners will be well positioned. By championing open standards, Upscale AI is giving infrastructure teams the architectural flexibility they need without compromising hyperscale performance.

‍

--------------------------------------------------------------------------------------------------------------------------------------------------

January 21, 2026

We are excited to announce that Premji Invest is co-leading Upscale AI's Series A, and to be partnering with founders Rajiv Khemani and Barun Kar, along with the rest of the Upscale AI team.

‍

We are witnessing a massive architectural shift within the data center. As we enter a multi-trillion dollar capex cycle, the ratio of network-spend to compute-spend is rising alongside cluster sizes. This creates an opportunity for Upscale AI to become the critical silicon provider, providing a step-change in performance and efficiency in a market that will be worth tens of billions in value. Members of the founding team previously architected Innovium, the only startup in the last decade to take significant market share from Broadcom (Premji led their Series E prior to their $1.1B acquisition by Marvell). We are thrilled to back this world-class team again as they revolutionize the AI networking market.

History & Problem Statement: the Rise of the AI Factory

The deployment of GenAI is driving one of the largest, fastest, and most coordinated capex cycles in modern history of technology. We are transitioning from the era of general-purpose cloud computing, characterized by virtualization of CPU-centric workloads, to the era of the "AI Factory." In this new paradigm, the fundamental unit of compute is not a single server but a tightly synchronized cluster of tens of thousands of accelerators, where massive parallelism requires low-latency coordination across devices.

‍

While the rise of LLMs recently resurfaced memory bottlenecks (driving demand for HBM), the continued scaling of accelerator compute has now pushed the bottleneck to I/O: the emerging "networking wall." As monolithic chip scaling approaches the reticle limit and Moore's Law slows, further scale can no longer be easily achieved by simply making individual accelerators larger or packing transistors more densely. Instead, improvements increasingly depend on efficiently distributing computation across multiple devices.

‍

In the CPU era, the challenge was instruction feeding. In the early deep learning era, the challenge was matrix multiplication throughput (FLOPS). Today, with mixture-of-experts (MoE) models and massive context windows, the challenge is data movement. Models grow 10x per year, requiring constant exchange of gradients and weights via model and tensor parallelism. GPUs often spend significant cycles idle, waiting for data to arrive from its neighbors. This "traffic jam" caps the ROIC for the entire AI industry.

Scale-up vs. scale-out

The industry distinguishes between two networking systems, "scale-up" and "scale-out". While these terms are used informally to describe communication within a single rack vs. across multiple racks, that physical distinction is becoming increasingly blurred. A clearer conceptual definition of a scale-up domain is one that shares memory semantics, i.e. a system that presents distributed accelerators as a single machine with a "shared pool" of memory. In contrast, scale-out networks rely on explicit data transfers and are more forgiving of latency.

In the near term, "rack-scale" competitiveness relies on performant scale-up silicon, a requirement shared by both hyperscalers and merchant XPU vendors. Furthermore, while the space is undoubtedly competitive, the recent proliferation of open networking standards reflects growing commercial appetite for high-performance and interoperable alternatives to vertically integrated solutions.

Upscale AI has emerged as the definitive answer to this mandate. Their SkyHammer (scale-up) [1.1]architecture represents flexible silicon purpose-built for AI workloads, combining the competitive performance traditionally associated with proprietary fabrics with the ecosystem benefits of open, standards-based protocols.

Opportunity

There is no merchant silicon provider today that offers a ground-up scale-up switch that is:

• Open: compatible with UALink (the AMD/Intel standard) 
• Optimized: free from Ethernet’s legacy baggage 
• Neutral: unaligned with a specific GPU vendor

‍

Enter Upscale AI.

‍

Firstly, Upscale AI is the only switch engineered to support both major open standards for scale-up: UALink (backed by AMD/Intel and others) and ESUN (Ethernet for scale-up networking, backed by AMD, Meta, MSFT and others). Upscale AI also supports UEC (Ultra Ethernet consortium, backed by AMD, Intel, Broadcom, and others).

‍

Secondly, Upscale AI's approach outperforms Ethernet. SkyHammer is a purpose-built ASIC designed from the ground up for scale-up networking. By avoiding legacy networking bloat (IP routing, legacy MACs), Upscale AI achieves strong "cut-through" latency, supporting deterministic flow control and adaptive load handling, while outperforming Broadcom's Ethernet adaptions. The technical moat is how SkyHammer handles memory semantics. In a scale-up pod, GPU A needs to read data from the memory of GPU B. NVLink does this natively. Ethernet does not: it wraps the data in a packet, sends it, unwraps it, and acknowledges it. Critically, SkyHammer reduces overhead to near zero for massive "all-to-all" bandwidth.

‍

Thirdly, Upscale AI is architecting the "Switzerland" of the AI data center. They hope to partner with all major clouds, working across any type of GPU (including custom ASICs at the large hyperscalers). Beyond silicon, Upscale AI is the 2nd largest contributor to SONiC, the open-source OS that ensures interoperability across different ASIC platforms (vendors provide their own SAI libraries that SONiC manipulates) and fast time-to-market. Their switches are also future-proofed for optical engines.

‍

Finally, Upscale's team has a stellar track record of taping out high-performance switches, a highly specialized area with a steep learning curve and little margin for error. Their founders and lead engineers bring rare, first-hand experience building leading-edge networking silicon and shipping reliably in complex production environments.

Conclusion

The "I/O wall" is now one of the defining challenges of the next decade. As AI workloads drive larger, more tightly coupled accelerator fabrics, networking has moved from a supporting role to a core architectural constraint. In response, the industry demands a dedicated, low-latency, memory-semantic fabric open to all.

‍

Upscale AI is a first-mover that has positioned itself at the center of the AI infrastructure shift. They provide the critical connectivity that allows companies to compete with established players on a rack-scale level. This neutrality is Upscale AI’s biggest strength: in a market eager to diversify away from proprietary solutions, Upscale AI represents a rare and high-value alternative. Combining this scarcity with a massive TAM and a superior open-standard approach, we are convinced this is the team destined to build a defining infrastructure company in the AI era.

‍

A doctor still reads every AI diagnosis. A lawyer still checks every AI brief. A CPA still signs every AI return. Not because the AI is too slow. Because when it's wrong in a high stakes domain, it cant be held responsible.

‍

So we keep the human in the loop. But what if we didn't have to? What if you could guarantee, with mathematical certainty, that the AI's answer is correct — and let the human go work on something higher-order than babysitting it? That's the premise of formal verification, and it's what Pramaana is building.

What is formal verification?

Formal verification isn't a new term. It's a sixty-year-old discipline that has quietly underwritten the things you already trust. Every modern CPU is formally verified before it tapes out, fly-by-wire aircraft software is certified against standards built on formal methods, and AWS uses TLA+ to verify the protocols behind S3 and DynamoDB.

‍

The concept is simple. You define a system in mathematical terms and write down a property you want to be true. A prover — Lean, Coq, Isabelle — mathematically checks whether the property holds across every possible input. This is completely different from vanilla testing — the output is a guarantee.

‍

The reason this discipline has lived inside math, code, and hardware for sixty years and nowhere else is two things. First, those domains hand you ground truth for free. A proof is right if it type-checks. Code is right if it matches the spec. Everything else — tax, contracts, medical guidelines, regulated finance — is written in English, contradicts itself in places, gets reinterpreted by courts and committees, and exists in no machine-readable form anywhere. There is no answer key.

‍

Second, even where you knew what you wanted to prove, the act of formalizing it was prohibitively expensive. Translating English statute into formal logic required someone fluent in both formal verification and the target domain, working at roughly a page a week. The entire US tax code at that rate would have taken a small army a decade. So nobody bothered.

What changed?

Two things happened, together but on separate tracks.

‍

The first is that AI made the process of formalizing a domain a lot easier than before. Pre-AI, it used to require a PhD doing it by hand, one statute at a time. AI, given it is an anything-to-anything converter in its most basic sense, can theoretically take a messy real-world domain and formalize it, assuming some guardrails.

‍

The second is that the ability to prove something also got easier. Lean is the language used to prove correctness and it draws from a library of math proofs called Mathlib. Mathlib is a community effort that has now become a real ecosystem. AI-driven proof generation is no longer speculative — it's something a serious research lab like Pramaana can build on.

What's the catch?

If this were easy, the foundation labs would have done it. They won't, not because they don’t want to, but because the chain of problems is long and definitely not straightforward.

‍

First, LLMs are notoriously bad at formalizing axioms in Lean directly. The training data is too thin and the type checker is too unforgiving. So you can't just point Claude at the Internal Revenue Code (in the case for tax) and ask for a formal encoding. You need an intermediate domain-specific language designed to be LLM-friendly and prover-friendly at the same time.

‍

Second, even if you translate the real, messy world into formal code, how do you ensure the translation is faithful? If two encoded rules contradict each other, is that a bug in your encoding or a real ambiguity in the statute? Pramaana's stack handles this through back-translation and an extreme amount of human checking.

‍

Third, how do you know what to formalize in the first place? Lets take tax as a first use case - it is not a closed system. It's open-ended, with judgment calls and jurisdictional variance and constant updates. To make it tractable, you have to define what counts as in-scope, draw boundaries, and ensure every situation maps to at least one rule. That scoping methodology is its own piece of IP — process knowledge learned over hundreds of formalizations.

How Pramaana works in practice

Pramaana is first and foremost a formal verification platform. Its first use case is tax. The company has been formalizing chunks of the US tax code — encoding statutes, regulations, and case law into machine-checkable rules. Think of it as the Mathlib of tax law. Every section formalized is permanent infrastructure that compounds. The tax code has 35,000 sections; they've started with the highest-impact ones and are working outward. The same template already scales to accounting rules like GAAP, banking rules like FDIC for financial institutions, medical prognosis and even football rules!

‍

The corpus is the foundation. The real product is what happens when you run a query against it. Imagine a mid-market business that just restructured, acquired something new, changed its sales practices. KPMG and Deloitte don't want the account. The CPA is doing ad-hoc work, not audit-proof analysis. The CFO wants to know: am I following the law, and am I saving the tax I'm supposed to be saving? Pramaana formalizes the question, runs it against the encoded rules, and returns an answer with a machine-checkable proof — or tells you exactly which rule blocks it and why. No iterative chat. No 95%. A guarantee.

Our thesis

The consensus AI thesis in 2026 so far is that the app layer is dead. The foundation labs will eat everything. We disagree, and Pramaana is the cleanest expression of our counter-thesis.

‍

The pattern is older than AI. In 2008, every venture investor said the same thing about the cloud: AWS will eat the app layer, but it didn't. Salesforce, Workday, Coupa, Anaplan, ServiceNow — all got built on top of the same substrate that was supposed to eat them. The app layer didn't die, it figured out where to add value the substrate couldn't.

‍

The same thing happens with AI. The model layer is the substrate. The companies that win at the app layer are the ones that can build full-stack initially - own the model, data, application surface - to show the art of the possible, while giving them optionality to provide the infrastructure for others to build other apps on top.

‍

Formal verification is one of those gaps. It started as a niche academic topic and is becoming infrastructure as AI moves into the high-stakes, autonomous use cases where 99% is unusable.

The team, and the signal

This is one of the rare teams that breaks our prior assumptions about how companies get built. Ranjan Rajagopalan led Google Maps Moderation — keeping a planetary-scale live database accurate. Krishnan Raghavan spent three years at Glean building the first version of Glean Assistant and fighting hallucinations, until he realized solving them is a research problem, not a product problem. Sanjay Ganapathy was a Staff Research Engineer at Google DeepMind with a decade in AI research — a core contributor to the Gemini models, builder of its tool-use system. All three are IIT Madras batchmates.

‍

The market signal has been hard to ignore. In their first year, Pramaana pulled 1,500+ registrations to their inaugural Verification Summit with mostly word of mouth. They assembled a world-class investor syndicate while residing outside the country. They got major press coverage on just seed raise. That's a lot for a team that's been around a year.

The shape of the bet

For sixty years, formal verification was the unsexy plumbing behind everything we already trusted. Pramaana is turning it into the engine of everything we couldn't trust before. We're excited to invest in their $27M round with friends at Khosla Ventures, Accel, Nexus, Boldcap and Unbound. Onwards.