Databricks vs Snowflake (7/7): The 2026 Scorecard — the Numbers, the Verdict, and Which Certification to Take

TL;DR — Databricks takes the week 3–1 with two draws, winning the architecture narrative, the table-format endgame and AI momentum. Snowflake won the one round that survives an audit: the money, and the fact that its numbers are filed with the SEC rather than announced from a stage. The most useful conclusion after seven days is that the rivalry made both platforms converge — so pick on workload fit, and learn the open formats they now both speak.

Six days ago we started with two founding stories. A Berkeley research project that became a company so its users would take the research seriously, and two Oracle kernel engineers who decided the only way to build a cloud warehouse was to stop retrofitting an on-prem one. Everything since — the benchmark war, the format war, the AI arms race — grew out of those two starting positions.

Today we close the ledger: the actual numbers as of August 2026, an honest guide to which one you should put on your own stack, and the certification comparison this series has been building toward.

The 2026 Scorecard

One caveat governs this entire table, and it matters more than any single cell. Snowflake's numbers are audited and filed with the SEC. Databricks' numbers are company-claimed. Databricks is private, has not filed an S-1, and discloses "annualized revenue run rate" at conferences and in funding press releases. Those are not the same species of number, and no honest scorecard should pretend otherwise.

Three of those rows deserve plain English, because they carry the whole argument.

Run rate is not revenue. An "annualized revenue run rate" takes a recent period — often a single month or quarter — and multiplies it out to a year. It is forward-looking arithmetic, not a result. Snowflake's $4.47B is money that actually arrived across twelve months and was checked by auditors. Databricks' $6.9B is what the last stretch implies a year would look like. Fast-growing companies naturally look larger under run-rate framing, which is exactly why private companies prefer it.

Net revenue retention (NRR) measures the same customers over time. Take everyone who spent money with you a year ago, ignore all new logos, and ask what that identical cohort spends now. 125% means last year's customers grew their spend by a quarter. 178% — Snowflake in FY2022 — means they nearly doubled it. The slide from 178% to 125% is not customers leaving; it's the expansion rate normalizing as the base gets huge and finance teams start policing consumption. Databricks' >140% is genuinely better, and also unaudited.

RPO is contracted future money. Remaining performance obligations is the value of signed commitments not yet recognized as revenue. Snowflake's $9.77B growing 42% is a stronger forward signal than the 29% current growth rate, which is part of why the market pays for it. Databricks publishes no equivalent, so this row simply cannot be compared.

The Two Stacks: Advantages and Disadvantages

Before the pick-one guidance, it's worth naming what each platform actually is, component by component, and being honest about what each one costs you.

The Databricks stack: Apache Spark as the distributed execution engine, with the Photon vectorized query engine underneath Databricks SQL; Delta Lake as the native table format (open-sourced April 2019, fully open at Delta Lake 2.0 in June 2022) with Apache Iceberg support native since June 2025 and the Tabular acquisition behind it; Unity Catalog for governance and lineage, open-sourced June 12, 2024; MLflow for experiment tracking and model lifecycle; Mosaic AI for model serving, vector search and Agent Bricks; Lakebase, the managed serverless Postgres built on the Neon acquisition; and a notebooks-first workflow in Python, Scala, SQL and R.

Advantages. One engine covers batch, streaming, SQL and ML training, so you are not stitching three products together. The ML and agent tooling is native rather than bolted on — MLflow, feature serving and Mosaic AI live inside the same governance boundary. Open formats are genuine here: Delta Lake sits under Linux Foundation governance and Unity Catalog is open-sourced, so the metadata layer is not a hostage. Multi-language support means data scientists and engineers share one workspace. Unstructured data, model artifacts and tables are governed by the same catalog.

Disadvantages. The operational burden is real: cluster configuration, autoscaling policy, runtime versions and job orchestration are all yours to get right, and getting them wrong is expensive rather than merely slow. The learning curve is the steepest part — Spark's execution model (shuffles, skew, caching, partition sizing) has to be understood, not just used. Cost predictability is the weakest point of the two: the DBU bill and the separately-billed cloud infrastructure arrive from different vendors, and reporting puts the infrastructure layer at roughly another 50–100% on top of DBUs. Notebook-first workflows drift toward untested, unreviewed code without deliberate engineering discipline. And the financial transparency gap cuts here too — there is no audited statement to underwrite a multi-year platform commitment.

The Snowflake stack: virtual warehouses as elastic, isolated compute over a shared storage layer, with micro-partitioned columnar storage and zone-map-style data skipping instead of traditional indexes (compressed files on object storage, pulled into compute on demand — the design described in the founders' SIGMOD 2016 paper); Snowpark for Python/Java/Scala pushdown; Iceberg tables plus the Polaris open catalog, donated to the Apache Software Foundation and graduated to a top-level Apache project on Feb 18, 2026; Cortex AI functions, Cortex Agents and Snowflake Intelligence (GA Nov 4, 2025); Arctic, its 480B-parameter Apache-2.0 open-weight MoE model; Streamlit for in-platform data apps; Snowflake Postgres from the Crunchy Data acquisition; and a SQL-first workflow throughout.

Advantages. Near-zero infrastructure management — no clusters to size, no runtimes to patch, no JVM tuning; a warehouse is a T-shirt size and an auto-suspend timer. Workload isolation is clean, because separate warehouses over the same data mean one team's bad query cannot starve another's dashboard. Time to first dashboard is the fastest in the category, and a SQL-fluent analyst is productive on day one with no distributed-systems knowledge. The credit price bundles the underlying cloud infrastructure, so there is one bill, not two. Catalog openness is credible: Polaris is a genuine Apache project, not a vendor-controlled repository. And you can audit the vendor's financials before you commit.

Disadvantages. Non-SQL work is second-class by design; Snowpark narrowed the gap but heavy ML training and complex streaming remain a stretch compared with a Spark-native platform. The abstraction that removes ops burden also removes control — when a query is slow, your tuning surface is clustering keys, warehouse size and query rewriting, not the execution plan itself. Costs are predictable per-credit but easy to waste through idle warehouses and oversized defaults. The proprietary storage engine underneath the platform is the real lock-in point, and Iceberg tables mitigate it only for the data you deliberately keep in Iceberg. The ecosystem outside SQL and BI is thinner. And the FY2022→FY2026 slide in net retention from 178% to 125% is the market's own signal that expansion inside existing accounts is getting harder.

Which One Should You Actually Pick?

On cost, the two are more similar than the marketing suggests, and both are consumption-priced, which is worth spelling out because it is where budgets die.

Consumption pricing means there is no flat licence fee. You are billed for resources used, metered by the second. Snowflake bills on three meters: compute credits (per-second, 60-second minimum, priced by warehouse size, edition, cloud and region), storage per compressed TB per month, and data transfer. Databricks bills in DBUs — Databricks Units, a normalized unit of compute capacity, also per-second, at a rate that varies by workload type. The structural difference that catches teams out: Databricks' DBU bill does not include the underlying cloud infrastructure. The AWS, Azure or GCP compute, storage and networking arrive on a separate bill from the cloud provider, and reporting on Databricks pricing puts that at roughly another 50–100% on top. Snowflake's credit price bundles the infrastructure.

Neither is inherently cheaper. A well-tuned warehouse with auto-suspend beats a badly sized all-purpose cluster; a job cluster running Spark with spot instances beats an oversized warehouse left running. What consumption pricing really means is that your architecture decisions show up on next month's invoice, which is also why both companies' growth rates are so sensitive to customer belt-tightening.

The Certification Question

Here the two platforms diverge more sharply than they do on technology.

Databricks: one price, flat ladder.

Every paid Databricks exam costs the same $200. Associate exams are 45–48 scored questions in 90 minutes; the Professional exams are 59 scored questions in 120 minutes. There are no formal prerequisites on any of them — Databricks recommends 6+ months of hands-on experience for Associate and 1+ years for Professional, but nothing gates registration. Renewal is simple and blunt: every certification expires after two years and you renew by re-sitting the current version of the exam. No continuing-education alternative surfaced in the research.

Snowflake: tiered ladder with a dependency chain.

Exams delivered in India get a 20% discount, confirmed on Snowflake's official page. Two version notes matter if you're reading older blog posts: Core is COF-C03, which replaced COF-C02 during 2026, and the Gen AI specialty is moving GES-C01 → GES-C02 — the GES-C01 page itself states it is being retired. The research surfaced a price discrepancy on the Gen AI specialty between the index page ($225) and the older cert page ($375), so check the current GES-C02 page before budgeting for it.

The structural difference is the renewal dependency chain. Every SnowPro Advanced exam requires an active SnowPro Core certification. Core is valid two years, renewable through continuing education or by passing an Advanced exam — which also extends Core by two years. Let Core lapse and your Advanced credential's standing is at risk with it. Databricks' flat two-year re-sit is more expensive in exam-hours; Snowflake's is cheaper if you keep climbing, and a trap if you stop.

Realistic path costs, list price:

• Databricks data engineer: Data Engineer Associate + Professional = $400
• Snowflake data engineer: Associate: Platform + Core + Advanced Data Engineer = $650
• Every paid Databricks credential: ~$1,400 at the official index snapshot
• Every Snowflake Advanced and Specialty track: well over $3,000 on top of Core

Two format details worth understanding. Snowflake's exams are commonly reported to use a scaled passing score of 750 out of 1000 — a scaled score is not "75% of questions correct." Raw scores are converted onto a fixed scale so that different exam forms of unequal difficulty are equally hard to pass, which means the number of questions you actually need right shifts between forms. The SnowPro Associate: Platform exam is also reported to be unproctored — no webcam, no lockdown software, no invigilator watching. That lowers friction considerably compared with a proctored exam — and it also tells you roughly how much weight the industry puts on that particular rung of the ladder.

Free paths first, on both sides. Databricks Academy's self-paced courses are free to all users, and Databricks Free Edition gives free hands-on platform access. Snowflake University offers free self-paced courses and free digital badges, with the Hands-On Essentials series free. On both platforms you can reach genuine working competence without paying anything; the exam fee buys the credential, not the knowledge.

On whether any of it is "worth it" in salary terms — we are not going to tell you, because the data does not exist. The research pass for this series went looking specifically for a primary source: a labour-statistics dataset, a named survey with published methodology, an actual dated community thread. It found none. Every salary premium, pass-rate percentage and "1-in-6 job postings" statistic circulating about these certifications traces back to sites that sell exam prep. Cert-prep vendors claim large salary lifts. That is a claim by parties with a direct financial interest, not a finding. The genuinely verifiable signal is much narrower and still useful: both vendors added dedicated generative-AI credentials in the 2025–2026 window — Databricks' Generative AI Engineer Associate plus the Context Engineer Associate that went GA on July 29, 2026, and Snowflake's Gen AI specialty already on its second exam version. That tells you where both companies believe the work is going.

Advice by role. If you're an analytics engineer at a Snowflake shop, take Core and stop unless your job actually needs an Advanced track. If you're a data engineer with Spark in your stack, the Databricks Associate is the cheapest credible signal at $200. If you're early-career and platform-agnostic, do the free accreditations and free hands-on tiers on both, then certify on whichever platform your target employers actually run. If you're already senior, certifications are a resume filter, not a differentiator — a public repository of real pipeline work outperforms both.

What to Learn From This

1. Read who audits the number before you read the number. "Run rate," "ARR," "annualized" and "bookings" are all legitimate terms that are also legitimate ways to look bigger than a filed revenue figure. When a private company's growth rate outruns a public competitor's by 50 points, part of that gap is real and part is measurement convention.
2. Always label the fiscal year. Snowflake's fiscal year ends January 31, so "FY2026" spans February 2025 to January 2026. Secondary coverage mislabels this constantly. When you compare two vendors' quarters, check that you are comparing the same calendar months.
3. Bet on formats, not vendors. The single most portable skill from this whole rivalry is Apache Iceberg, which both platforms now read and write natively after a decade of fighting about it. Skills tied to an open format survive a migration; skills tied to a proprietary surface do not.
4. Cost architecture is a design decision, not a procurement one. Under consumption pricing, cluster sizing, auto-suspend, file layout and partitioning are line items on the invoice. Learn the billing model of whatever platform you run as carefully as you learn its SQL dialect.
5. Sequence certifications around renewal mechanics, not prestige. Snowflake's Advanced tier requires active Core, so the ladder wants continuous motion. Databricks resets everyone every two years at a flat $200. Price the five-year cost of staying current, not the one-time exam fee.
6. When you cannot find a primary source, say so. The cert-salary question in this article was unanswerable with the evidence available, and writing "cert-prep vendors claim X" is more useful to a reader than laundering a marketing number into a fact. Apply the same standard to vendor benchmarks — as Day 4 showed, that's exactly how a benchmark war starts.

Final Verdict

Final tally: Databricks 3 — Snowflake 1, with 2 draws.

Databricks won architecture (Day 2), the format war (Day 5) and the AI pivot (Day 6). What it actually won in each case was narrative control: the lakehouse framing that made "warehouse versus lake" feel like a solved problem, the Tabular acquisition that bought the Iceberg founders during Snowflake's own conference week, and an AI story compelling enough to carry a private valuation from $43B in September 2023 to $188B by July 2026.

Snowflake won the money (Day 3), and that win is more durable than a single point suggests. It is the only one of the two with a share price, audited statements, a $9.77B contracted backlog and a decade of numbers anyone can check. It also won something the scoreboard has no column for: simplicity. A team can run Snowflake without a distributed-systems specialist on staff, and that has always been the product's real argument.

The draws — origins and the 2021 benchmark war — are the honest verdicts. Neither company invented the other's category, and neither won a benchmark fight that no neutral party ever adjudicated.

The closing thought is the one the whole week points at: the rivalry made both platforms converge. Databricks built a warehouse and a Postgres; Snowflake built agents and a Postgres. Both open-sourced a catalog. Both speak Iceberg. Both signed nine-figure deals with a frontier model provider. The lakehouse-versus-warehouse question that defined 2021 is close to obsolete in 2026, and what's left is the boring, correct answer: pick on workload fit, price the consumption model honestly, and invest your own learning in the open formats neither company can take away from you.

Which one did you pick, and did the numbers above change your mind? Argue it out in the comments — the scorecard was ours, the verdict is yours.

Sources

• Snowflake Q4 and full-year FY2026 results
• Databricks: 65% YoY growth, $5.4B revenue run-rate (Feb 2026)
• CNBC: Databricks revenue growth tops 80% to $6.9B annualized (June 2026)
• TechCrunch: Databricks hits $188B valuation (July 2026)
• Snowflake (SNOW) market cap history
• Databricks certification catalogue
• Snowflake SnowPro certification catalogue
• Databricks acquires Neon for serverless Postgres (May 2025)
• Snowflake 10-Q disclosing $164.5M Crunchy Data cash consideration
• Snowflake Intelligence GA and the $200M Anthropic agreement