Alembic Technologies Raises $145M to Build World's Fastest Private Supercomputer for Causal AI

Alembic melted GPUs chasing causal A.I. -- now it's running one of the fastest supercomputers in the world

Alembic Technologies has raised $145 million in Series B and growth funding at a valuation 13 times higher than its previous round, betting that the next competitive advantage in artificial intelligence will come not from better language models but from proprietary data and causal reasoning. The San Francisco-based startup, which builds AI systems that identify cause-and-effect relationships rather than mere correlations, is using a significant portion of the capital to deploy what it claims is one of the fastest privately owned supercomputers ever built -- an Nvidia NVL72 superPOD that will power its enterprise-grade causal AI models. The investment, led by Prysm Capital and Accenture with participation from Silver Lake Waterman, Liquid 2 Ventures, NextEquity, Friends & Family Capital and WndrCo, positions Alembic among a select group of well-funded AI laboratories transforming how corporations make multimillion-dollar decisions. The funding round and the company's strategic direction reflect a broader shift taking place in enterprise AI as the performance gap between competing large language models narrows. While startups and tech giants have poured billions into building ever-larger chatbots, Alembic is pursuing a different thesis: that the real value in AI will accrue to systems that can process private corporate data to answer questions that generic models cannot. "As powerful artificial intelligence models increasingly converge in capability, the key competitive advantage shifts to proprietary data," said Tomás Puig, Alembic's founder and chief executive, in an interview with VentureBeat. "Getting a real edge isn't about using the best LLM; it's leveraging the unique information rivals can't access." Puig illustrated the problem facing enterprise executives: "Imagine I run a CPG company and I install the latest ChatGPT. I ask, 'Hey, ChatGPT, give me a strategy for how to increase my revenue share in the northeast.' Then your competitor down the road asks the exact same question. How much trouble are you in when they get the exact same answer?" How a broke startup on Mac Pros discovered a breakthrough that changed everything The dramatic valuation increase -- from roughly $50 million at the Series A to approximately $645 million now, according to people familiar with the matter -- reflects a fundamental transformation in Alembic's technology and market positioning since its previous funding round. When the company raised its Series A in early 2024, it was primarily a signal processing and correlation analytics company focused on marketing measurement. "Causal did not exist as a technology for us till after the Series A," Puig told VentureBeat. The company was so resource-constrained that it couldn't even run simulations to test whether its causal models would work. The breakthrough came after the Series A when the company finally had enough capital to test its theories. "We were so broke that we couldn't even run the simulation to see if it worked," Puig recalled. When they did run the tests -- initially on an "army of Mac Pros" because they didn't yet have GPU infrastructure -- they discovered something unexpected: their causal model worked not just for marketing analytics but across virtually any business domain with time-series data. "We started adding capabilities as customers requested them, which was just sensible -- iterative," Puig explained. "We found out the model works across a huge majority of data universally. What we thought might be a model for a specific vertical ended up being a full, generalized foundational model." That discovery transformed Alembic from a marketing technology vendor into a company building what Puig describes as "the entire central nervous system of the enterprise across all verticals -- not just sales, marketing, supply chain, finance, and beyond." Why cause-and-effect AI matters more than correlation for enterprise decision-making Causal AI is a fundamentally different approach from the correlation-based analytics that dominate most business intelligence tools and even many AI systems. Where traditional analytics might show that social media engagement correlates with sales increases, causal AI can determine whether the social media activity actually caused the sales lift -- or whether both were driven by some third factor, like a viral news event. The distinction matters enormously for executives making budget allocation decisions. "Most businesses are not short on data," Puig said. "They are short on answers." For Alembic's customers, which now include Delta Air Lines, Mars, Nvidia and several Fortune 500 companies across financial services, technology and consumer packaged goods, the platform can answer previously unanswerable questions about marketing effectiveness, operational efficiency and strategic investments. "Alembic's ability to connect marketing exposure directly to business outcomes -- with speed, precision and granularity -- is what made this relationship so transformative for us," said Alicia Tillman, chief marketing officer at Delta Air Lines. "Unlike traditional measurement tools, Alembic gave us a unified view across channels and campaigns, unlocking insights we simply couldn't access before." The airline used Alembic to quantify the revenue lift from its Team USA Olympics sponsorship within days of activation, directly linking brand activities to ticket sales -- a type of measurement that has eluded marketers for decades. Traditional attribution models either ignore brand-building entirely or assign it vague "awareness" metrics that don't translate to financial impact. "It's very transformative," Puig said of the customer impact. "What's interesting is that executives themselves are the users of our software and our outputs. It's not a tool used by a single campaign manager." Inside the two-story liquid-cooled supercomputer that literally melted GPUs Alembic's decision to invest heavily in owned computing infrastructure rather than rely on cloud providers stems from both the technical demands of its causal models and the extreme data sensitivity requirements of its enterprise customers. The company is deploying an Nvidia NVL72 superPOD -- a massive liquid-cooled system equipped with Nvidia's most advanced Blackwell graphics processing units -- in partnership with data center operator Equinix in San Jose, Calif. According to Puig, Nvidia informed Alembic that it is the only non-Fortune 500 company in the world to operate such a system. The need for this level of compute stems from how Alembic's models work. Unlike large language models that are trained once on historical data and then deployed, Alembic's system uses "online and evolving" models built on spiking neural networks -- brain-inspired architectures that continuously learn as new data arrives. "It creates itself as you feed it data, like human evolution," Puig explained. "The model is singular, but it ends up creating a different brain for every single company." This continuous learning happens at massive scale. When a customer brings in data, Alembic's system automatically permutates through billions of possible combinations of how that data could be analyzed -- testing every conceivable way to slice metrics and dimensions to find the strongest causal signals. That level of computation requires what Puig calls "F1 car" infrastructure rather than the "production Porsche" offered by cloud providers. The company writes custom CUDA code and low-level GPU kernels optimized specifically for causal inference workloads -- optimizations that aren't possible on standard cloud configurations. The approach has proven so demanding that Alembic famously once melted down its GPUs by pushing them beyond their thermal limits. "We literally just drive these circuits so hard that we need the F1 car version and we have to have access to it," Puig said. The move to liquid-cooled systems addresses that problem, but it also enables Alembic to run workloads that would cost orders of magnitude more on cloud platforms. "We did the math -- if we were to buy just one subsection of our compute from AWS, it would be $62 million a year," Puig said. Owning the infrastructure costs "a fraction of that." The supercomputer strategy serves another crucial purpose: data sovereignty. Many of Alembic's customers -- particularly in financial services, consumer packaged goods and regulated industries -- have contractual prohibitions against putting sensitive data on Amazon Web Services, Microsoft Azure or Google Cloud. "CPG companies do not want any data to exist on Amazon, ever," Puig said. "They simply won't allow it. Some customers refuse to use Microsoft, others avoid different providers. And certain banks and financial institutions are legally prohibited from using cloud platforms at all." By operating its own infrastructure in neutral data centers, Alembic can serve customers who would never consider cloud-based analytics -- a competitive moat that would be difficult for hyperscale cloud providers to replicate. How Jensen Huang read a news article and changed Alembic's destiny Alembic's relationship with Nvidia illustrates both the startup's technical ambitions and how the chip giant supports promising AI companies. Nvidia is Alembic's founding enterprise customer, exclusive supercomputing partner and a key technical collaborator -- though notably not an investor. The relationship began in an unlikely way. After Alembic announced its Series A funding in early 2024, Nvidia co-founder and CEO Jensen Huang read the VentureBeat coverage and emailed his staff suggesting they explore the company, according to Puig. Because Alembic didn't yet have a contact form on its website, an Nvidia director reached out via LinkedIn. The partnership nearly foundered on a basic constraint: computing capacity. After Alembic delivered its first causal analysis -- which took weeks to generate on an array of Mac Pros -- Nvidia asked if they could produce weekly reports. "I said no, because it took weeks on this army of machines," Puig recalled. When Alembic said they could do it with GPUs but couldn't secure the necessary compute -- cloud providers at the time required committee approvals and offered two- to six-week lead times with no guarantees -- Nvidia intervened directly. The chip maker arranged for Equinix to provide a private cage in Northern Virginia with sufficient power capacity and helped Alembic source its first H100 GPU cluster. "Without that, the company would never have existed," Puig said. "We couldn't get the compute in the configuration we needed anywhere else." The partnership has since deepened. Alembic uses Nvidia's AI Enterprise software suite, including specialized libraries like cuGraph for graph processing and TensorRT for high-speed inference. The tight integration, Puig said, allows "our research teams to leverage multi-exaflop-level compute and Nvidia's algorithmic software stack. This integration is one of our secret weapons: we spend more time on breakthrough research and mathematics and less time on repetitive low-level engineering." Nvidia's support extended beyond technology. When Alembic kept destroying GPUs under extreme workloads -- pushing chips so hard that thermal stress cracked circuit boards -- Nvidia fast-tracked the startup's access to next-generation liquid-cooled systems. "The funny reason we got [the NVL72]," Puig said, "is because when we melted the chips, Nvidia was literally annoyed with how often they had to service our warranty." From Olympics sponsorships to viral candy moments: How Fortune 500s measure what was unmeasurable Alembic's customer roster has expanded rapidly as enterprises seek ways to measure AI and marketing investments that traditional analytics cannot capture. The company now works with Delta Air Lines, Mars, multiple Fortune 500 technology and financial services firms, and Texas A&M University's athletics program. The use cases span far beyond Alembic's original marketing focus. Mars used the platform to measure the sales impact of changing candy shapes for themed promotions. A Fortune 500 technology company expanded its sales pipeline by 37% using Alembic's attribution models. Financial services firms are using it to connect CEO public appearances and co-marketing expenditures to actual fund flows. "Alembic helped us move past impression counts to show what actually drove net-new investment," said the head of co-marketing at a Fortune 200 financial services company. "For the first time, we could see how our CEO in the public eye and our co-marketing dollars with exchanges translated into real fund flows." For Mars, the ability to measure previously unmeasurable activities has transformed decision-making. "We are using math to liberate creativity," said Gülen Bengi, lead global chief marketing officer for Mars and global chief growth officer for Mars Snacking. "Our fans and communities create billions of organic conversations and content about our brands. When a viral moment happens, we normally know it's directionally positive, but we can't attribute the sales uplift or its place in the customer journey. Alembic's Causal AI is a breakthrough, allowing us to move beyond correlation to see exactly how that organic conversation created a sequence that directly impacted sales." The platform can predict revenue, close rates and customer acquisition up to two years in advance with 95% confidence, according to Puig. "What they were doing before was they actually literally did not know about certain things," he said, describing how customers previously estimated the value of stadium naming rights or major sponsorships without ever measuring actual dollar impact. "Now you can go and be like it had this effect on this much P&L, and this is where it's flowing, and you can know within days or near real time." Why Google, Meta and Nielsen can't easily replicate what Alembic built Alembic operates in a competitive landscape that includes traditional marketing measurement vendors like Nielsen, analytics platforms from Google and Meta, and emerging AI-powered analytics startups. But Puig argues the company has built structural advantages that would be difficult to replicate. First, the company's causal models rely on proprietary mathematics developed over years and protected by patents. "You would have to start from scratch," Puig said. "This is not like an LLM that uses a transformer that has a paper, and you could attempt to recreate. You'd actually have to go and recreate the methodology from scratch." Second, the massive computing requirements create a natural barrier. Alembic operates at "foundational model levels of compute, not like even something you would run from [AWS] Sagemaker," Puig said. "We're talking about hundreds of millions of dollars a year" in equivalent cloud costs. Third, the data sovereignty requirements of enterprise customers create opportunities for neutral third parties that hyperscale cloud providers struggle to address. As one venture capital investor noted, enterprises increasingly worry about putting strategic data into systems owned by potential competitors. Finally, Alembic's ability to work with messy, fragmented data reflects years of engineering that preceded its causal AI breakthrough. "The first four [or] five years of the company's life was building that giant signal processor that dealt with messy data," Puig said. "We would not be able to do it if we had not taken all that time." Why Alembic's contrarian bet on private data could reshape enterprise AI The $145 million funding round validates a contrarian bet in an AI landscape dominated by the race to build ever-larger language models. While OpenAI, Anthropic and others compete on whose chatbot can write better code or answer more trivia questions, Alembic is building infrastructure for a different kind of intelligence -- one that understands cause and effect in the messy, proprietary data that defines each company's unique competitive position. The company's evolution from a bootstrapped startup running simulations on Mac Pros to operating one of the world's fastest private supercomputers mirrors the broader maturation of enterprise AI. As the technology moves from experimentation to mission-critical deployment, companies need more than general-purpose models trained on public data. They need systems that can process their private information to answer questions their competitors cannot. Puig's thesis -- that private data becomes the key differentiator as public models converge -- resonates with how other technologies evolved. Search engines commoditized access to public information, making proprietary data more valuable. Cloud computing made infrastructure a utility, elevating the importance of what you build on top of it. If large language models similarly converge in capability, the competitive advantage flows to whoever can best extract intelligence from data others cannot access. The company is already testing its technology beyond marketing analytics. Pilots are underway in robotics, where causal models could help autonomous systems understand how actions lead to outcomes. New product lines are launching, including the GPU-accelerated database that customers are buying separately. The ambition, Puig said, is to become "the central nervous system" of the enterprise -- the layer that connects cause and effect across every business function. Whether Alembic can deliver on that vision remains to be seen. The company operates in complex enterprise environments where sales cycles are long and integration challenges are significant. Competitors aren't standing still, and the technical moats that protect it today may erode as causal AI techniques become better understood. But for now, Alembic occupies a unique position. It has marquee customers achieving measurable results. It has infrastructure that would cost hundreds of millions to replicate on cloud platforms. It has proprietary mathematics refined over years of dealing with messy enterprise data. And it has $145 million to scale what Puig describes as a fundamental shift from correlation to causation. In his interview with VentureBeat, Puig drew a parallel to quantitative hedge funds that use mathematics to gain trading advantages that general-purpose AI cannot match. "ChatGPT still can't equal Renaissance Technologies," he said, referring to the secretive firm that has generated historic returns through quantitative models. The comparison captures Alembic's core insight: that in a world where everyone has access to the same general-purpose AI, sustainable advantage comes from specialized systems that understand the cause-and-effect relationships hiding in your data. It's a bet that the future of enterprise AI looks less like a universal chatbot and more like a private intelligence engine -- one that, to Puig's original point, prevents your competitor from getting the same answer when they ask the same question.

Alembic Technologies raises $145M and buys an Nvidia-powered supercomputer to accelerate 'causal AI' - SiliconANGLE

Alembic Technologies raises $145M and buys an Nvidia-powered supercomputer to accelerate 'causal AI' Causal artificial intelligence startup Alembic Technologies Inc. said today it has raised $145 million in a Series B growth round that increases its valuation almost 16-fold. It's using a big chunk of those funds to invest in what it says is one of the fastest privately owned supercomputers ever built - a new Nvidia Corp. NVL72 superPOD that'll serve as the foundation of its enterprise-grade AI models. Alembic claims to be a pioneer of enterprise-focused causal AI models that focus on generating marketing intelligence. Causal AI models are designed to understand and model cause-and-effect relationships, as opposed to just identifying patterns and correlations. Alembic said it's leveraging causal AI to generate richer insights into the underlying factors behind certain consumer behaviors and events, so it can enhance its customers' decision-making. In other words, Alembic is trying to develop more reliable AI models that can safely be used in enterprise environments, and to do that it needs to leverage its customers' all-important private data. The startup said its Causal Engine acts like a private intelligence layer for AI models and creates a compounding flywheel. By surfacing better insights, companies can create superior products and strategies for their customers, which then help to capture more, potentially even richer data, allowing further insights to be generated. Alembic co-founder and Chief Executive Tomás Puig said there's really not much difference among the most powerful large language models today in terms of their performance or capabilities. It's the proprietary data those models can access that provides the main competitive advantage. "Getting a real edge isn't about using the best LLM," he explained. "It's about leveraging the unique information your rivals can't access. The exclusive data a company generates protects its strategy from generic model outputs, where two competitors could get the same answers." The size of today's round indicates that Alembic's new supercomputer doesn't come cheap, but Puig stressed that it's a vital investment for the company. It's planning to deploy the NVL72 superPOD, which is equipped with Nvidia's most powerful Blackwell graphics processing units, in a vast DGX AI supercomputing cluster in partnership with the data center colocation provider Equinix Inc. The cluster will run the Nvidia AI Enterprise software suite to provide a "high-performance computing backbone" for its causal AI models. Puig said the system has been customized to run its continuous-learning spiking neural networks and spatio-temporal graph construction algorithms, enabling it to scale to support its growing customer base and continue generating fresh causal insights in real time. It's notable that this is the second supercomputing cluster Alembic has acquired, as part of its strategic focus to ensure bicoastal redundancy and avoid being locked in to any cloud vendors. By investing in a dedicated, private AI fleet, Alembic benefits from guaranteed resource availability and gains the ability to better optimize its stack for causal AI workloads. According to Puig, Alembic needs access to the world's most advanced and capable computing platforms in order to be able to advance casual AI further. "[NVL72] allows our research teams to fully leverage multi-exaflop-level compute and Nvidia's algorithmic software stack," he explained. "This deep integration is one of our secret weapons - we spend more time on breakthrough research and mathematics and less time on repetitive low-level engineering, enabling us to deliver Causal answers to partners like Accenture in near real-time." Today's round was co-led by Prysm Capital and Accenture Plc, and saw participation from Liquid 2 Ventures, NextEquity, Friends & Family Capital, WndrCo and SLW. Accenture Chair and CEO Julie Sweet said causal AI is critical for enterprise adoption of AI, because regulated companies need both high performance and extreme reliability. "Alembic's Causal AI provides exactly that by moving the enterprise beyond correlation to deliver the verifiable, cause-and-effect insights leaders need to act with decisive speed," she said.