Monday, March 5, 2026, 10:00 a.m. Pacific. Somewhere inside a Samsung research lab or on an HP infrastructure floor, a dashboard goes dark. No maintenance window, no deprecation warning with a grace period—just an HTTP 503, and a few hours later, confirmation that every metric, every gradient log, every meticulously tagged hyperparameter run is now permanently deleted. Neptune.ai’s hosted service had reached its OpenAI-mandated end of life, and with it evaporated the experiment history of dozens of enterprise teams that hadn’t finished migrating. That same morning, over inside CoreWeave’s cloud console, a Weights & Biases user was probably dragging a slider through a hyperparameter sweep comparison chart, oblivious to the carnage on the other side of the MLOps fence. This is the 2026 experiment-tracking comparison nobody planned for. It’s not the usual side-by-side feature matrix—it’s an autopsy of one platform, an ongoing physical exam of the other, and a messy, involuntary migration story for everyone caught in between.

The Acquisitions That Rewired the Category In a matter of months, both major commercial experiment-tracking platforms got bought. The outcomes couldn’t have diverged more starkly. CoreWeave picked up Weights & Biases in mid-2025 in a deal pegged somewhere between $1.4 billion and $1.7 billion, depending on which SEC watcher you trust. CoreWeave, itself a Nvidia-backed cloud heavyweight valued north of $23 billion at the time, treated W&B as the white-collar piece of its “end-to-end AI developer stack.” The platform kept its name, kept its multi-cloud posture, and kept onboarding logos—AstraZeneca, Nvidia, Toyota, and over 1,400 other organizations. Then OpenAI struck Neptune.ai in December 2025 for less than $400 million in stock. Neptune’s 60-ish employees mostly got invites to join the mothership in San Francisco. Its Warsaw-born codebase got folded into OpenAI’s internal training infrastructure, with chief scientist Jakub Pachocki talking up “fully traceable, auditable iterations” for future foundation model runs. That’s wonderful for GPT-whatever-comes-next. For the rest of the world—Samsung, Roche, HP, InstaDeep—the outcome was a hard cutoff: no more SaaS, no more self-hosted image downloads after March 8, zero possibility of recovering data left behind. Community reactions on Hacker News captured the frustration succinctly. One developer wrote, “i moved from wb to neptune because of wb's slow ui, i loved neptune so much i'm seriously considering building a ‘clone’.” Another posted, “Weights & Biases if they are serious and need performance and scale,” as if acknowledging the sudden narrowing of the commercial playing field. Two former competitors. One is now a cloud provider’s software tentacle. The other is a proprietary internal utility that you and I will never touch again.

Where the Architectures Actually Differed Before the shutdown, you would have compared these tools on philosophy. W&B was the team-first, collaboration-obsessed workhorse built around wandb.init() and wandb.log(). Its polish showed in the dashboards, the mobile app, the reporting layer that let managers gawk at loss curves without writing a line of Python. Neptune, meanwhile, chased a different beast: extreme-scale metadata fidelity. Neptune’s true differentiator was its ability to ingest over one million data points per second without downsampling, then let you filter 100,000-plus runs with sub-second latency. Per-layer gradient tracking, activation logging, and run forking from any checkpoint—those were first-class citizens, not afterthoughts. That architecture made it a favorite for teams debugging training instabilities at scales where W&B’s UI could, by multiple user accounts, get sluggish. W&B countered with a broader surface area: built-in hyperparameter sweeps, an artifact lineage system, collaborative Reports, and the W&B Model Registry that ties experiments to deployment artifacts. After the CoreWeave deal, the company bolted on ARIA, an AI research agent that sifts through thousands of runs to propose fixes to underperforming hyperparameter configs. No equivalent existed on Neptune’s public roadmap before the lights went out. The integration story followed the same split. Neptune’s one-line integration auto-detected 22-plus frameworks, making it a quick spin-up for polyglot teams. W&B demanded more manual wiring but gave deeper hooks into its ecosystem—Weave for LLM observability, direct CI/CD coupling to its registry. For self-hosted diehards, Neptune had an ace: full on-prem deployment for compliance-heavy shops. W&B’s on-prem option is limited and gated behind enterprise contracts. In a post-Neptune world, that’s suddenly a real gap, not just a footnote in an RFP.

The Money Side: Per-Seat vs. Per-Byte Pricing always split buyers. W&B went the Salesforce route—$50 to $60 per user per month for early-stage teams, with custom enterprise pricing that tacks on SSO, HIPAA, customer-managed encryption keys, and a Bring Your Own Bucket storage option. It’s a seat model, plain and simple. Viewer seats, Full seats, ingestion metering for Weave—each variable adds a line to the bill. Neptune charged per data volume: a Team plan at $50 per user per month, but a Lab plan at $250 per user per month with a 10-billion-data-point ceiling and 10 TB of storage. Overage? Two dollars per gig per month. If your team logged millions of metric points per run, Neptune got expensive. But for metadata-heavy workloads, the economics occasionally inverted; some teams reported W&B bills that crept to $180,000 a year while MLflow could handle 80% of the job for near-zero licensing cost. That dynamic leads to a slightly uncomfortable truth about 2026 buying behavior. As one community analyst put it, “Most teams choose an experiment tracker based on whatever the first ML engineer installed.” That inertia routinely outweighs formal TCO analysis—until a shutdown or a ballooning invoice forces the issue.

The Involuntary Migration Surge Neptune’s rush job of a sunset—roughly three months from acquisition to data erasure—created a forced migration wave that nobody talks about in polite vendor keynotes. The company published an exporter tool that dumped runs into Parquet files and a loader script to push them into W&B. A technical webinar walked teams through mapping Neptune’s project hierarchy to W&B’s organization, teams, projects, runs model. But for large installations, the real cost wasn’t the script; it was rewriting the API plumbing, restructuring data models, and allocating local storage to stage multi-terabyte exports before reimporting. Some teams went the other way—towards MLflow, Comet, ZenML, or more niche alternatives like Minfx.ai. MLflow’s 26,000 GitHub stars and deep cloud integrations (SageMaker Unified Studio, Databricks) made it a natural lifeboat, especially for teams that already had a Databricks footprint and valued open-source governance. Comet, fresh off a $4.5 million infusion for its agentic AI focus, pitched itself as the “team-friendly” alternative with an open-source LLM observation path. ZenML—pipeline-centric, with SOC 2 and BYOK support—appealed to organizations that wanted infrastructure abstraction without rebuilding everything when the next vendor vanishes. The broader lesson isn’t just about tooling; it’s about architectural neutrality. AWS’s May 2026 move to natively embed MLflow into SageMaker experiments underscored the trend: cloud vendors are betting that open-source standards win the lock-in wars. Enterprises are starting to listen.

The Landscape Beyond Feature Checklists What’s more interesting than a feature matrix is the gravitational pull reshaping the entire category. Experiment tracking used to be a standalone discipline. In 2026, it’s becoming a feature inside three larger movements. First, LLM observability has swallowed a chunk of the attention—and the venture dollars. W&B’s Weave product sits squarely in this space, letting teams trace, evaluate, and guardrail generative AI outputs alongside traditional experiment metrics. Langfuse, Arize AI, and Braintrust are all carving out niches, while Pure experiment-tracking startups are nearly absent from recent funding rounds. The money is chasing “failure memory” for AI agents and formal verification of LLM outputs, not run-logging dashboards. Second, the idea of a self-improving loop—where every production failure becomes labeled training data for the next deployment—is starting to make a static dashboard look quaint. For LLM-native teams, platforms that close the generate-simulate-evaluate-optimize cycle may soon matter more than who has the best hyperparameter scatter plot. This doesn’t kill the need for experiment lineage; it just changes what you demand from the tool. W&B is clearly aware; ARIA is a step towards automated analysis, not just passive visualization. Third, vendor neutrality is no longer a philosophical debate. It’s a procurement checklist item. When two independent platforms disappear into two AI giants within one year, the “too big to fail” argument evaporates. For all of CoreWeave’s multi-cloud reassurances, W&B now lives inside a vertically-integrated GPU cloud company. That’s not the same as being a Switzerland; it’s more like being Luxembourg.

What Buyers Should Actually Do Right Now If you’re still stuck on a Neptune self-hosted instance, the math is simple: you’re running unmaintained software with no security patches and no image repository. Your migration clock isn’t ticking—it already expired. The only question is whether you can reconstruct your experiment history before a compliance auditor asks for it. If you’re evaluating new platforms, the conversation likely collapses to three paths. For most commercial teams already bought into the W&B ecosystem, staying put makes sense—the integrations, the ARIA agent, the polished UI, and a corporate parent that’s publicly traded and incentivized to keep the lights on. The caveat is cost at scale; negotiate seat tiers aggressively and pin down storage pricing early. If compliance or cost pushes you towards self-hosting, MLflow remains the default. It’s not as slick, it won’t auto-suggest hyperparameter fixes, and you’ll spend more time building dashboards. But it’s free, it’s open, it’s inside every major cloud, and it won’t get acquired. The wildcard is the LLM observability layer. Even if you pick a traditional tracker for core experiments, you’ll probably end up running a second tool—Langfuse, Arize, or Weave—for production genAI traces. Getting those two systems to talk to each other without duct tape is the real engineering challenge nobody’s solved out of the box. As one Hacker News commenter put it, a little mournfully, “i loved neptune so much i'm seriously considering building a ‘clone’.” The sentiment captures something bigger than nostalgia for a tool. It’s a market that’s rapidly consolidating around a few big players, leaving very little room for the kind of focused, fast, metadata-first engineering that Neptune represented. Whether that’s a natural maturation or a loss of real innovation depends on what gets built next—and who’s allowed to keep building it.