Alembic Applied sciences has raised $145 million in Sequence B and progress funding at a valuation 13 instances greater than its earlier spherical, betting that the subsequent aggressive benefit in synthetic intelligence will come not from higher language fashions however from proprietary information and causal reasoning.
The San Francisco-based startup, which builds AI programs that establish cause-and-effect relationships somewhat than mere correlations, is utilizing a good portion of the capital to deploy what it claims is without doubt one of the quickest privately owned supercomputers ever constructed — an Nvidia NVL72 superPOD that may energy its enterprise-grade causal AI fashions.
The funding, led by Prysm Capital and Accenture with participation from Silver Lake Waterman, Liquid 2 Ventures, NextEquity, Associates & Household Capital and WndrCo, positions Alembic amongst a choose group of well-funded AI laboratories reworking how companies make multimillion-dollar choices.
The funding spherical and the corporate's strategic path replicate a broader shift going down in enterprise AI because the efficiency hole between competing massive language fashions narrows. Whereas startups and tech giants have poured billions into constructing ever-larger chatbots, Alembic is pursuing a distinct thesis: that the actual worth in AI will accrue to programs that may course of non-public company information to reply questions that generic fashions can’t.
"As highly effective synthetic intelligence fashions more and more converge in functionality, the important thing aggressive benefit shifts to proprietary information," mentioned Tomás Puig, Alembic's founder and chief govt, in an interview with VentureBeat. "Getting an actual edge isn't about utilizing the very best LLM; it's leveraging the distinctive info rivals can't entry."
Puig illustrated the issue going through enterprise executives: "Think about I run a CPG firm and I set up the newest ChatGPT. I ask, 'Hey, ChatGPT, give me a method for methods to enhance my income share within the northeast.' Then your competitor down the street asks the very same query. How a lot hassle are you in once they get the very same reply?"
How a broke startup on Mac Professionals found a breakthrough that modified every thing
The dramatic valuation enhance—from roughly $50 million on the Sequence A to roughly $645 million now, in keeping with individuals conversant in the matter — displays a basic transformation in Alembic's expertise and market positioning since its earlier funding spherical.
When the corporate raised its Sequence A in early 2024, it was primarily a sign processing and correlation analytics firm centered on advertising and marketing measurement. "Causal didn’t exist as a expertise for us until after the Sequence A," Puig instructed VentureBeat. The corporate was so resource-constrained that it couldn't even run simulations to check whether or not its causal fashions would work.
The breakthrough got here after the Sequence A when the corporate lastly had sufficient capital to check its theories. "We have been so broke that we couldn't even run the simulation to see if it labored," Puig recalled. Once they did run the checks — initially on an "military of Mac Professionals" as a result of they didn't but have GPU infrastructure — they found one thing surprising: their causal mannequin labored not only for advertising and marketing analytics however throughout nearly any enterprise area with time-series information.
"We began including capabilities as prospects requested them, which was simply smart—iterative," Puig defined. "We discovered the mannequin works throughout an enormous majority of information universally. What we thought may be a mannequin for a selected vertical ended up being a full, generalized foundational mannequin."
That discovery remodeled Alembic from a advertising and marketing expertise vendor into an organization constructing what Puig describes as "your entire central nervous system of the enterprise throughout all verticals — not simply gross sales, advertising and marketing, provide chain, finance, and past."
Why cause-and-effect AI issues greater than correlation for enterprise decision-making
Causal AI is a essentially totally different method from the correlation-based analytics that dominate most enterprise intelligence instruments and even many AI programs. The place conventional analytics may present that social media engagement correlates with gross sales will increase, causal AI can decide whether or not the social media exercise really induced the gross sales raise — or whether or not each have been pushed by some third issue, like a viral information occasion.
The excellence issues enormously for executives making price range allocation choices. "Most companies will not be brief on information," Puig mentioned. "They’re brief on solutions."
For Alembic's prospects, which now embrace Delta Air Strains, Mars, Nvidia and a number of other Fortune 500 firms throughout monetary companies, expertise and client packaged items, the platform can reply beforehand unanswerable questions on advertising and marketing effectiveness, operational effectivity and strategic investments.
"Alembic's means to attach advertising and marketing publicity on to enterprise outcomes—with velocity, precision and granularity—is what made this relationship so transformative for us," mentioned Alicia Tillman, chief advertising and marketing officer at Delta Air Strains. "Not like conventional measurement instruments, Alembic gave us a unified view throughout channels and campaigns, unlocking insights we merely couldn't entry earlier than."
The airline used Alembic to quantify the income raise from its Staff USA Olympics sponsorship inside days of activation, straight linking model actions to ticket gross sales—a kind of measurement that has eluded entrepreneurs for many years. Conventional attribution fashions both ignore brand-building solely or assign it imprecise "consciousness" metrics that don't translate to monetary influence.
"It's very transformative," Puig mentioned of the shopper influence. "What's attention-grabbing is that executives themselves are the customers of our software program and our outputs. It's not a device utilized by a single marketing campaign supervisor."
Contained in the two-story liquid-cooled supercomputer that actually melted GPUs
Alembic's choice to speculate closely in owned computing infrastructure somewhat than depend on cloud suppliers stems from each the technical calls for of its causal fashions and the acute information sensitivity necessities of its enterprise prospects.
The corporate is deploying an Nvidia NVL72 superPOD — a large liquid-cooled system geared up with Nvidia's most superior Blackwell graphics processing items — in partnership with information heart operator Equinix in San Jose, Calif. In response to Puig, Nvidia knowledgeable Alembic that it’s the solely non-Fortune 500 firm on the earth to function such a system.
The necessity for this degree of compute stems from how Alembic's fashions work. Not like massive language fashions which are educated as soon as on historic information after which deployed, Alembic's system makes use of "on-line and evolving" fashions constructed on spiking neural networks — brain-inspired architectures that repeatedly be taught as new information arrives.
"It creates itself as you feed it information, like human evolution," Puig defined. "The mannequin is singular, nevertheless it finally ends up creating a distinct mind for each single firm."
This steady studying occurs at huge scale. When a buyer brings in information, Alembic's system mechanically permutates by means of billions of attainable mixtures of how that information could possibly be analyzed — testing each conceivable strategy to slice metrics and dimensions to search out the strongest causal indicators. That degree of computation requires what Puig calls "F1 automobile" infrastructure somewhat than the "manufacturing Porsche" provided by cloud suppliers.
The corporate writes customized CUDA code and low-level GPU kernels optimized particularly for causal inference workloads — optimizations that aren't attainable on commonplace cloud configurations. The method has confirmed so demanding that Alembic famously as soon as melted down its GPUs by pushing them past their thermal limits. "We actually simply drive these circuits so laborious that we’d like the F1 automobile model and we have now to have entry to it," Puig mentioned.
The transfer to liquid-cooled programs addresses that drawback, nevertheless it additionally permits Alembic to run workloads that will price orders of magnitude extra on cloud platforms. "We did the mathematics—if we have been to purchase only one subsection of our compute from AWS, it could be $62 million a 12 months," Puig mentioned. Proudly owning the infrastructure prices "a fraction of that."
The supercomputer technique serves one other essential objective: information sovereignty. A lot of Alembic's prospects — significantly in monetary companies, client packaged items and controlled industries — have contractual prohibitions in opposition to placing delicate information on Amazon Net Companies, Microsoft Azure or Google Cloud.
"CPG firms don’t want any information to exist on Amazon, ever," Puig mentioned. "They merely received't enable it. Some prospects refuse to make use of Microsoft, others keep away from totally different suppliers. And sure banks and monetary establishments are legally prohibited from utilizing cloud platforms in any respect."
By working its personal infrastructure in impartial information facilities, Alembic can serve prospects who would by no means contemplate cloud-based analytics — a aggressive moat that will be tough for hyperscale cloud suppliers to copy.
How Jensen Huang learn a information article and altered Alembic's future
Alembic's relationship with Nvidia illustrates each the startup's technical ambitions and the way the chip large helps promising AI firms. Nvidia is Alembic's founding enterprise buyer, unique supercomputing companion and a key technical collaborator — although notably not an investor.
The connection started in an unlikely manner. After Alembic introduced its Sequence A funding in early 2024, Nvidia co-founder and CEO Jensen Huang learn the VentureBeat protection and emailed his employees suggesting they discover the corporate, in keeping with Puig. As a result of Alembic didn't but have a contact type on its web site, an Nvidia director reached out by way of LinkedIn.
The partnership almost foundered on a primary constraint: computing capability. After Alembic delivered its first causal evaluation — which took weeks to generate on an array of Mac Professionals — Nvidia requested if they might produce weekly experiences. "I mentioned no, as a result of it took weeks on this military of machines," Puig recalled.
When Alembic mentioned they might do it with GPUs however couldn't safe the mandatory compute — cloud suppliers on the time required committee approvals and provided two- to six-week lead instances with no ensures — Nvidia intervened straight. The chip maker organized for Equinix to supply a non-public cage in Northern Virginia with enough energy capability and helped Alembic supply its first H100 GPU cluster.
"With out that, the corporate would by no means have existed," Puig mentioned. "We couldn't get the compute within the configuration we wanted wherever else."
The partnership has since deepened. Alembic makes use of Nvidia's AI Enterprise software program suite, together with specialised libraries like cuGraph for graph processing and TensorRT for high-speed inference. The tight integration, Puig mentioned, permits "our analysis groups to leverage multi-exaflop-level compute and Nvidia's algorithmic software program stack. This integration is one in every of our secret weapons: we spend extra time on breakthrough analysis and arithmetic and fewer time on repetitive low-level engineering."
Nvidia's assist prolonged past expertise. When Alembic stored destroying GPUs below excessive workloads — pushing chips so laborious that thermal stress cracked circuit boards — Nvidia fast-tracked the startup's entry to next-generation liquid-cooled programs. "The humorous motive we obtained [the NVL72]," Puig mentioned, "is as a result of once we melted the chips, Nvidia was actually irritated with how usually they needed to service our guarantee."
From Olympics sponsorships to viral sweet moments: How Fortune 500s measure what was unmeasurable
Alembic's buyer roster has expanded quickly as enterprises search methods to measure AI and advertising and marketing investments that conventional analytics can’t seize. The corporate now works with Delta Air Strains, Mars, a number of Fortune 500 expertise and monetary companies companies, and Texas A&M College's athletics program.
The use instances span far past Alembic's unique advertising and marketing focus. Mars used the platform to measure the gross sales influence of adjusting sweet shapes for themed promotions. A Fortune 500 expertise firm expanded its gross sales pipeline by 37% utilizing Alembic's attribution fashions. Monetary companies companies are utilizing it to attach CEO public appearances and co-marketing expenditures to precise fund flows.
"Alembic helped us transfer previous impression counts to indicate what really drove net-new funding," mentioned the pinnacle of co-marketing at a Fortune 200 monetary companies firm. "For the primary time, we might see how our CEO within the public eye and our co-marketing {dollars} with exchanges translated into actual fund flows."
For Mars, the flexibility to measure beforehand unmeasurable actions has remodeled decision-making. "We’re utilizing math to liberate creativity," mentioned Gülen Bengi, lead international chief advertising and marketing officer for Mars and international chief progress officer for Mars Snacking. "Our followers and communities create billions of natural conversations and content material about our manufacturers. When a viral second occurs, we usually realize it's directionally constructive, however we are able to't attribute the gross sales uplift or its place within the buyer journey. Alembic's Causal AI is a breakthrough, permitting us to maneuver past correlation to see precisely how that natural dialog created a sequence that straight impacted gross sales."
The platform can predict income, shut charges and buyer acquisition as much as two years upfront with 95% confidence, in keeping with Puig. "What they have been doing earlier than was they really actually didn’t learn about sure issues," he mentioned, describing how prospects beforehand estimated the worth of stadium naming rights or main sponsorships with out ever measuring precise greenback influence. "Now you may go and be prefer it had this impact on this a lot P&L, and that is the place it's flowing, and you may know inside days or close to actual time."
Why Google, Meta and Nielsen can't simply replicate what Alembic constructed
Alembic operates in a aggressive panorama that features conventional advertising and marketing measurement distributors like Nielsen, analytics platforms from Google and Meta, and rising AI-powered analytics startups. However Puig argues the corporate has constructed structural benefits that will be tough to copy.
First, the corporate's causal fashions depend on proprietary arithmetic developed over years and guarded by patents. "You would need to begin from scratch," Puig mentioned. "This isn’t like an LLM that makes use of a transformer that has a paper, and you possibly can try and recreate. You'd really should go and recreate the methodology from scratch."
Second, the huge computing necessities create a pure barrier. Alembic operates at "foundational mannequin ranges of compute, not like even one thing you’ll run from [AWS] Sagemaker," Puig mentioned. "We're speaking about tons of of tens of millions of {dollars} a 12 months" in equal cloud prices.
Third, the info sovereignty necessities of enterprise prospects create alternatives for impartial third events that hyperscale cloud suppliers wrestle to deal with. As one enterprise capital investor famous, enterprises more and more fear about placing strategic information into programs owned by potential opponents.
Lastly, Alembic's means to work with messy, fragmented information displays years of engineering that preceded its causal AI breakthrough. "The primary 4 [or] 5 years of the corporate's life was constructing that enormous sign processor that handled messy information," Puig mentioned. "We might not be capable to do it if we had not taken all that point."
Why Alembic's contrarian guess on non-public information might reshape enterprise AI
The $145 million funding spherical validates a contrarian guess in an AI panorama dominated by the race to construct ever-larger language fashions. Whereas OpenAI, Anthropic and others compete on whose chatbot can write higher code or reply extra trivia questions, Alembic is constructing infrastructure for a distinct sort of intelligence — one which understands trigger and impact within the messy, proprietary information that defines every firm's distinctive aggressive place.
The corporate's evolution from a bootstrapped startup working simulations on Mac Professionals to working one of many world's quickest non-public supercomputers mirrors the broader maturation of enterprise AI. Because the expertise strikes from experimentation to mission-critical deployment, firms want greater than general-purpose fashions educated on public information. They want programs that may course of their non-public info to reply questions their opponents can’t.
Puig's thesis — that personal information turns into the important thing differentiator as public fashions converge — resonates with how different applied sciences advanced. Serps commoditized entry to public info, making proprietary information extra priceless. Cloud computing made infrastructure a utility, elevating the significance of what you construct on high of it. If massive language fashions equally converge in functionality, the aggressive benefit flows to whoever can greatest extract intelligence from information others can’t entry.
The corporate is already testing its expertise past advertising and marketing analytics. Pilots are underway in robotics, the place causal fashions might assist autonomous programs perceive how actions result in outcomes. New product strains are launching, together with the GPU-accelerated database that prospects are shopping for individually. The ambition, Puig mentioned, is to change into "the central nervous system" of the enterprise — the layer that connects trigger and impact throughout each enterprise perform.
Whether or not Alembic can ship on that imaginative and prescient stays to be seen. The corporate operates in complicated enterprise environments the place gross sales cycles are lengthy and integration challenges are vital. Opponents aren't standing nonetheless, and the technical moats that shield it in the present day might erode as causal AI strategies change into higher understood.
However for now, Alembic occupies a novel place. It has marquee prospects attaining measurable outcomes. It has infrastructure that will price tons of of tens of millions to copy on cloud platforms. It has proprietary arithmetic refined over years of coping with messy enterprise information. And it has $145 million to scale what Puig describes as a basic shift from correlation to causation.
In his interview with VentureBeat, Puig drew a parallel to quantitative hedge funds that use arithmetic to realize buying and selling benefits that general-purpose AI can’t match. "ChatGPT nonetheless can't equal Renaissance Applied sciences," he mentioned, referring to the secretive agency that has generated historic returns by means of quantitative fashions.
The comparability captures Alembic's core perception: that in a world the place everybody has entry to the identical general-purpose AI, sustainable benefit comes from specialised programs that perceive the cause-and-effect relationships hiding in your information. It's a guess that the way forward for enterprise AI seems much less like a common chatbot and extra like a non-public intelligence engine — one which, to Puig's unique level, prevents your competitor from getting the identical reply once they ask the identical query.
