Nebuly Raises $21M to Answer AI's Hardest Question: What Did It Return?

A $21 million bet on the receipt

Nebuly says it has raised $21 million in Series A funding to help enterprises measure the return on their AI initiatives. The round was led by Planven, the company says, with participation from Exor Ventures, Tenacity Capital, TGVG Capital, and Fastweb+Vodafone, and support from CDP Venture Capital SGR through its Artificial Intelligence Fund and Digital Transition Fund. Nebuly says it measures the return enterprises get from their AI agents, and that its platform connects to agents including Claude, ChatGPT, and Microsoft Copilot. The company says it has secured multiple six-figure contracts with large enterprises in telecommunications, automotive, industrial manufacturing, luxury, healthcare, and technology, and names Oura, Iveco Group, Fastweb+Vodafone, CNH, and Zeiss as customers. m:…5134

The funding lands as spending outruns measurement. Nebuly cites a Gartner survey finding that 85 percent of functional leaders plan to increase their AI spending this year, while only 22 percent of organizations have successfully scaled AI across multiple business units. The company also says Uber hit its annual AI budget four months into 2026, and that an Amazon project ran 860 percent over budget due to unfinished agent tasks. Those are Nebuly's examples, not independently verified figures, but they describe the gap Nebuly is selling into: money going out faster than anyone can say what it returned. m:…5134

Return needs a baseline, not a benchmark

Broadcom makes a parallel point from infrastructure. In a post on VCF migration outcomes, Broadcom's Justin Kaplan argues that completing a migration does not by itself prove value, and that leaders should validate performance against pre-migration baselines. The post cites Forrester's 2026 Total Economic Impact study of VMware Cloud Foundation 9, which projected a three-year ROI of 31 percent to 137 percent for a composite organization, with a medium-impact scenario of 73 percent. Kaplan's framework, which he calls VCF Migration Assurance, connects evidence across baseline, transition, and validation phases, using AppNeta for network measurements and synthetic application tests and DX NetOps for network performance and fault context, while Clarity by Broadcom relates the findings to business commitments. The post warns that financial ROI calculations must account for migration services, licensing, monitoring tools, operational overhead, and infrastructure upgrades. m:…0383

A small model that proved the point

Toloka offers a concrete case of measurement done the hard way. The company says an autoresearch agent trained 62 versions of a small model for bank support, and that the post-trained 4-billion-parameter model achieved 93.8 percent accuracy on unseen synthetic conversations. Training data consisted of 2,048 synthetic conversations, the company says, and a single training run took about an hour on one H100 GPU. Reinforcement learning after full fine-tuning improved performance on changed-fact evaluation datasets, according to Toloka, which compared full fine-tuning with LoRA for model training. m:…5282

The cost comparison is where the return argument sharpens. Toloka says the post-trained 4B model cost one fifty-eighth of the serving cost of Gemini 3.8 Flash, and was 22 times faster. But the company is candid about the ceiling: the model's performance was limited by underrepresented situations in the training data, and dataset limitations, specifically underrepresented edge cases like four-day transfers, remain the primary bottleneck. The training method worked; the data decided where it stopped. m:…5282

What the funding means

Taken together, the three posts describe a shift in what buyers are being asked to pay for. Nebuly is selling visibility into agent performance, costs, and revenue impact. Broadcom's Kaplan is selling a discipline: baselines before migration, evidence after, and costs counted in full. Toloka is selling proof that a small, measured model can beat a large one on cost and speed for a narrow job, and proof that the data, not the training method, decides where it stops. For investors, the question Nebuly's round raises is whether measuring AI return becomes a product category of its own, or a feature every platform is expected to include. m:…5134 m:…0383 m:…5282