TL;DR — Both companies bought an AI team in 2023, shipped an open-weight LLM in H1 2024, open-sourced an Iceberg catalog in June 2024, and bought a Postgres company within three weeks of each other in 2025. The difference is timing and coherence: Databricks paid ~$1.3B for MosaicML and compounded it into models, agents and a $1B+ AI revenue run-rate, while Snowflake spent 2024 swapping CEOs and answering for ~165 breached customer accounts. Cortex is a real catch-up — it just started later.
The format war ended in a shared standard, which meant the fight had to move somewhere neither side could copy the other's homework in a quarter. It moved to AI. And the strange thing about the 2023–2026 record is how closely the two roadmaps rhyme: same year for the AI acquisition, same half-year for the open model, same month for the open catalog, same three-week window for the Postgres buy.
Rhyming roadmaps are not the same as equal outcomes. What separates these two is not whether they saw the pivot — both did, early — but what happened to the thing they bought after they bought it.
The concepts you need
Four ideas carry the argument. None require you to have touched either platform.
What buying a foundation-model team buys. Not the model — models age out in months. You buy training infrastructure: the orchestration, checkpointing and GPU scheduling that keeps a thousand-GPU run alive at hour 400 instead of dying at hour 40, plus the engineers who know why runs die. The model depreciates; the ability to train the next one compounds.
Mixture-of-experts (MoE). A dense model pushes every token through every parameter. An MoE splits the feed-forward layers into many "experts" with a router that picks a few per token — you store all parameters but compute a fraction. Hence DBRX at 132B total / 36B active, Arctic at 480B / 17B. Memory tracks total; compute tracks active.
"AI functions in the warehouse." Cortex's core idea: call the model from SQL — SELECT AI_CLASSIFY(review_text) FROM reviews — inside the query engine you already have. The point is inheritance, not convenience: the call runs as your database role and respects the grants already on that table.
What an agent needs beyond a model. Tools (actions it may call), governance (whose identity it acts under, enforced outside the prompt), memory (state between steps). That is the tell for the 2025 acquisitions below: both companies shopped for catalogs, access control and databases, not models.
2023: two acquisitions, two very different assets
In May 2023 Snowflake acquired Neeva, the consumer search startup founded by ex-Google ad chief Sridhar Ramaswamy. Snowflake's own SEC filing puts the price at $185.4 million in cash — worth stating precisely, because earlier reporting circulated a looser "~$150M" figure before the filing landed. Ramaswamy joined as SVP of AI.
One month later, on June 26, 2023, Databricks announced a definitive agreement to acquire MosaicML for approximately $1.3 billion, inclusive of retention. MosaicML brought the MPT open-source model family and, more importantly, a training team — people whose job was making large models cheap to train on someone else's data.
Seven times the price for something that plugged directly into the product. Snowflake bought a search team and, without knowing it, a CEO.
February 2024: the CEO swap
On February 27, 2024, Frank Slootman retired as Snowflake CEO, staying on as non-executive Chairman. Ramaswamy — nine months into the building — was named CEO. It was, on paper, an AI-forward succession: put the AI guy in charge of the AI pivot.
Markets read it alongside the guidance. The announcement landed with a weak FY2025 forecast (~22% revenue growth against a Street expectation near 30%), and SNOW fell roughly 20–22%, from about $230 to about $179 — call it $17B of market cap in a day. (Our sources date the drop to Feb 28 and Feb 29, 2024 respectively; the mechanics are the same either way.)
That is the asymmetry that defines the year. Databricks entered 2024 integrating a training lab. Snowflake entered 2024 onboarding a new chief executive into a decelerating growth story.
H1 2024: DBRX vs Arctic
Both shipped an open-weight enterprise LLM within a month of each other, and both were mixture-of-experts.
| DBRX (Databricks) | Arctic (Snowflake) | |
|---|---|---|
| Announced | March 27, 2024 | April 24, 2024 |
| Total params | 132B | 480B |
| Active params | 36B (4 of 16 experts) | 17B (top-2 of 128 fine-grained experts) |
| Claim | Beat LLaMA2-70B, Mixtral, Grok-1; best GPT-3.5 | Enterprise-grade, Apache 2.0, open weights + code |
| Training cost | ~$10M (reported) | Under $2M, under 3 months (company-stated) |
| Built by | Mosaic Research — i.e. the team they bought | Snowflake AI Research |
Arctic is the more interesting engineering artifact — 128 fine-grained experts, 17B active, and a company claim of sub-$2M training cost that, if accurate, is a genuinely striking efficiency result. But neither model became the product. What mattered was which one seeded a durable capability, and Mosaic Research kept shipping into the platform long after DBRX stopped being newsworthy.
Six weeks later both open-sourced a catalog: Snowflake's Polaris (announced June 3, 2024, Apache 2.0, donated to the ASF around August 2024, now Apache Polaris) and Databricks' Unity Catalog OSS (June 12, 2024, backed by AWS, Google Cloud, Microsoft, NVIDIA and Salesforce, later hosted at LF AI & Data). Same week, Databricks announced Tabular — the company founded by Iceberg's original Netflix creators — during Snowflake Summit. Price was never disclosed; press estimates ran from $1B to $2B, so treat any specific number as reported, not confirmed. Ghodsi acknowledged a bidding war that included Snowflake.
May 2024: the breach that wasn't a platform breach
On May 30, 2024, Snowflake, Mandiant and CrowdStrike disclosed the UNC5537 campaign. The precision matters: this was not a compromise of Snowflake's core platform. Attackers used customer credentials harvested by infostealer malware to log into roughly 165 customer accounts that lacked MFA. Confirmed victims included Ticketmaster/Live Nation, AT&T (~110M customers' call and text metadata, disclosed July 12, 2024), Santander, Advance Auto Parts, Neiman Marcus and Pure Storage. Two suspects were later charged; one was arrested in Canada on a US extradition request in October 2024.
Concept: credential stuffing without MFA. Infostealer malware scrapes saved passwords off a laptop; the credentials get sold; the attacker simply logs in. No exploit, no zero-day — a valid password is the front door working as designed. MFA breaks the chain by making a stolen password insufficient on its own. Which is how "the platform wasn't breached" and "165 customers were" are both true.
Technically exonerating. Commercially expensive anyway. "Your customers didn't turn on MFA" is a true sentence that does not survive a headline cycle, and the fix — pushing MFA from optional to enforced — is an admission that optional was a product decision, not a law of nature. SNOW fell ~5% on the initial disclosure days and was down ~43% year-to-date by late August 2024, though the honest reading is that the breach was one pressure among several; the guidance cut was doing independent damage. Anyone attributing the 2024 chart to the breach alone is selling a cleaner story than the data supports.
2025: the Postgres pincer
Then the symmetry got almost comic. Databricks announced Neon — open-source serverless Postgres — for ~$1B on May 14, 2025, citing telemetry that 80% of databases provisioned on Neon were created automatically by AI agents. Three weeks later, on June 1–2, Snowflake announced Crunchy Data. Press reported "about $250 million"; Snowflake's own 10-Q states $164.5M in cash, "primarily for its talent and developed technology." Use the filing.
Neon became Lakebase (public preview June 11, 2025; GA on AWS later that year). Crunchy became Snowflake Postgres. Both bets rest on the same thesis: agents need a transactional database they can spin up and throw away, and the warehouse is the wrong shape for that.
Concept: OLTP versus the warehouse. A warehouse is OLAP — built to scan enormous column ranges, optimised for throughput, awful at thousands of tiny single-row reads and writes per second. OLTP (the Postgres world) is the inverse: row-oriented, millisecond point lookups, real transactions. An agent's working state — conversation history, task queues, tool-call results — is exactly that stream of small operations, and running it on a warehouse is slow and, under consumption pricing, expensive. Hence a disposable OLTP database beside the analytical one.
June 2026: the HTAP dream, and who got closest
Concept: HTAP. Since Gartner coined the term in 2014, HTAP — Hybrid Transactional/Analytical Processing — has been the database industry's white whale: one system that handles both the small fast transactional writes AND the huge analytical scans, so you stop copying data from the OLTP database into the warehouse through fragile ETL/CDC pipelines. Everyone has claimed it; nobody has fully delivered it, because the two workloads want physically opposite storage layouts (rows vs columns).
Snowflake's run at it came first: Unistore, announced in 2022, whose Hybrid Tables — row-oriented tables living inside the warehouse, queryable alongside analytical data — reached GA on AWS on October 30, 2024 (Azure followed October 2025). It is real and shipping, but it is HTAP inside the warehouse: the transactional workload comes to Snowflake's engine, on Snowflake's terms, at warehouse pricing.
Databricks' answer landed at the Data + AI Summit in June 2026, and it deliberately did not use the H-word. LTAP — "Lake Transactional/Analytical Processing" — is Lakebase (the Postgres it got from Neon) rearchitected so that transactional writes are mirrored into open columnar format on lake storage, which the analytical engines then read directly. The pitch: transactions and analytics on one copy of the data, in an open format, with no CDC pipeline in between — unification at the storage layer rather than inside a single do-both engine. Databricks says Lakebase already serves thousands of customers (Block, Zillow, Superhuman among the named ones) and handles about 12 million database launches a day — company-claimed numbers, no independent audit.
Keep the salt shaker handy, though. The Register's headline captured the honest caveat: Databricks "unifies OLTP and OLAP, depending on what counts as a copy" — a Postgres row store mirrored into columnar files is arguably still two physical representations; what has genuinely been removed is the pipeline you used to build and babysit between them. And the announcement is months old, not battle-proven. The fair scoring as of August 2026: Snowflake shipped the narrower, warehouse-shaped version first and has almost two years of GA behind it; Databricks just shipped the more ambitious, open-format version of the idea — the closest anyone has come to the HTAP dream at lake scale — but it is new enough that production war stories don't exist yet.
The head-to-head: AI-era moves, 2023–2026
| Databricks | Snowflake | |
|---|---|---|
| AI acquisition | MosaicML, ~$1.3B, June 2023 | Neeva, $185.4M, May 2023 |
| Open model | DBRX, 132B MoE, Mar 2024 | Arctic, 480B MoE, Apr 2024 |
| AI product line | Mosaic AI / Agent Bricks (beta June 11, 2025) | Cortex AI Functions + Cortex Agents (GA Nov 4, 2025) |
| Open catalog | Unity Catalog OSS, June 12, 2024 | Polaris, June 3, 2024 → Apache Polaris |
| Postgres buy | Neon, ~$1B, May 14, 2025 → Lakebase | Crunchy Data, $164.5M SEC, June 2025 → Snowflake Postgres |
| Other buys | Tabular (undisclosed, ~$1B+ reported, 2024), BladeBridge (2025), Tecton (2025, reported), Mooncake Labs (2025, reported), Panther Labs (announced June 2026, closed Aug 2026), Antimatter + SiftD.ai | Select Star (Nov 24, 2025), Datometry (Nov 2025, reported), Natoma (May 2026, reported) |
| Agent story | Agent Bricks: describe the task, platform auto-generates evals and optimizes quality/cost; Lakebase as agent-native OLTP | Snowflake Intelligence GA Nov 4, 2025 — 1,000+ customers deployed 15,000+ agents in the prior three months |
| Model partnership | OpenAI, Sept 25, 2025, ≥$100M committed spend (Anthropic deal from Mar 2025 unaffected) | Anthropic, Nov 4, 2025, $200M multi-year |
| Leadership | Ghodsi continuous throughout | Slootman → Ramaswamy, Feb 27, 2024 |
| AI revenue | >$1B AI run-rate (2025); $1.4B and ~26% of a $5.4B run-rate by Feb 9, 2026 | Not separately disclosed |
| Valuation arc | $62B (Dec 2024) → >$100B (Aug 2025) → $134B (Dec 2025) → $188B (July 2026, single-source) | Public; Q3 FY2026 revenue $1.21B, +29% YoY |
Choosing between the two AI stacks
Winning the round is not the same as being right for your team. Scoped strictly to AI and agents:
| Databricks Mosaic AI | Snowflake Cortex | |
|---|---|---|
| Advantages | Depth across the chain — pretrain, fine-tune, evaluate, serve, and run agents against an agent-native OLTP database inside one governance boundary. Owns the model layer instead of renting it; Mosaic Research kept shipping into the platform long after DBRX left the news cycle. Agent Bricks generates the evals from a task description — the step most teams skip and then regret. Model-neutral at the top (OpenAI ≥$100M committed spend alongside the earlier Anthropic deal). | Accessibility as architecture, not consolation: AI as a SQL function is usable by the analysts you already employ, putting "idea" and "working thing" days apart instead of a quarter. Governance is inherited, not rebuilt — the call obeys the grants already on that table. And the catch-up is real: Cortex AI Functions, Cortex Agents and Snowflake Intelligence all GA'd Nov 4, 2025. |
| Disadvantages | Surface area is the price of depth — one production agent touches compute config, Unity Catalog, vector search, an eval pipeline and a Postgres, each a thing to learn, operate and pay for. Premium skills: an ML-engineer stack, not an analyst stack. Agent Bricks was still beta in June 2025, and fast-moving surfaces mean rework. You control the compute, so you own the runaway-cost failure mode. | Started roughly eighteen months later, and the gap shows in disclosure: Databricks reports a $1.4B AI run-rate, Snowflake doesn't break AI revenue out at all. Thinner training story — Arctic was efficient (under $2M, company-stated) but nothing plays Mosaic Research's role of feeding the product continuously. The SQL surface is a ceiling as well as a floor. Agent governance is bought in (Select Star, Natoma) rather than long-owned. And the MFA default was a product decision buyers still remember. |
Short version: Databricks is the deeper stack if you have engineers to train and operate custom models; Cortex ships governed AI faster, with fewer people.
What to Learn From This
- Judge an acquisition by what it plugs into on day 30. MosaicML dropped into the product as a training capability; Neeva gave Snowflake a CEO — a leadership asset, not a product one. Ask which of a vendor's acquisitions changed the platform and which changed the org chart.
- Read the filing, not the headline. Neeva circulated at ~$150M and filed at $185.4M; Crunchy Data was reported "about $250 million" and filed at $164.5M. Filings correct press estimates in both directions.
- Budget an MoE model on active parameters. Serving cost tracks params computed per token (36B for DBRX, 17B for Arctic); memory tracks the total. Size inference off the headline count and you sized the wrong thing.
- Audit MFA enforcement, not MFA availability. UNC5537 needed no exploit — just valid credentials where the second factor was optional. The question is never "do we support MFA," it is "can any account still authenticate without it."
- Plan transactional state next to your lakehouse, not inside it. Agent memory is an OLTP workload. Both vendors bought a Postgres company in the same three weeks for that reason — design for it before your session store becomes a line item on a consumption bill.
- Resist single-cause explanations. SNOW's 2024 decline stacked a breach, a CEO transition and a guidance miss into two quarters; any account crediting one of the three is cleaner than the evidence.
Scorecard
Round 6: Databricks. The bet was earlier, seven times bigger, and — decisively — it compounded. MosaicML became Mosaic Research, which became DBRX, which became the model-serving and evaluation machinery under Agent Bricks; by February 2026 that stack was a disclosed $1.4B AI run-rate inside a $5.4B business. Snowflake's 2023 bet bought talent that became a CEO, which is not nothing, but it meant 2024 went to rebuilding leadership and trust rather than shipping. Credit where due: Cortex is genuine catch-up — Snowflake Intelligence GA'd in November 2025 with 15,000+ agents deployed in its first three months and a $200M Anthropic partnership behind it, and Arctic was the more efficient model on paper. It just started eighteen months behind, against a company that never had to look away.
Running tally: Databricks 3 — Snowflake 1 (2 draws).
Tomorrow
The finale — the 2026 scorecard, which platform to actually pick, and which certification is worth your money.
Sources
- Databricks signs definitive agreement to acquire MosaicML — $1.3B, June 2023
- Sridhar Ramaswamy named CEO of Snowflake — Feb 2024
- Announcing DBRX — 132B MoE, March 2024
- Snowflake Arctic, 480B open enterprise LLM — April 2024
- Snowflake data breach (UNC5537) overview — ~165 customer accounts, no MFA
- Databricks agrees to acquire Neon — ~$1B, May 2025
- Snowflake 10-Q (SEC) — Crunchy Data at $164.5M cash
- Snowflake Intelligence brings agentic AI to the enterprise — GA Nov 4, 2025
- Snowflake's Unistore Hybrid Tables reach general availability — GA on AWS Oct 30, 2024
- Databricks launches LTAP — Lake Transactional/Analytical Processing — Data + AI Summit, June 2026
- The Register on LTAP: "depending on what counts as a copy" — July 3, 2026
Top comments (0)