It was 20 below zero outside, but they were there because they were going to be part of the DNA of how the trucking industry works," said Tim Graczewski, Global Head of Confluent for Startups. He was describing a January interview with Hemut's founders, who joined the call from the office of a small trucking company in Poplar Grove, Illinois, where they had embedded themselves to learn how freight works.
Hemut builds an AI operating system for trucking. According to a Confluent case study, founder and CEO Loki Cheema learned the trade as a teenager at a motel front desk in Everett, Washington, helping long-haul truckers with bills of lading, logbooks and settlement statements, before taking degrees in computer science and linguistics at UCLA.
"It's a regulated industry that is running on archaic 1980s technology," Graczewski said. "You can't just go to this industry and give them something completely new. It has to be part of how they already think, and the regulations they work within."
Graczewski interviewed the founders with Sean Falconer, whom he described as Confluent's head of AI. According to Confluent, Hemut joined the third cohort of its Data Streaming AI Accelerator in February 2026 and had its real-time data foundation in production by March. Confluent says the platform has since processed more than 286 million events without losing one, and that one large deployment identified about 37,000 hours of manual work a year, worth roughly $1.3 million in labour costs. Hemut had already graduated from Y Combinator and raised a $1.8 million seed round at a $30 million valuation.
Capital is pooling in fewer, larger rounds
Global venture funding reached $159 billion in the third quarter of 2026, according to Crunchbase, down 25% from the previous quarter but up 53% from a year earlier. Crunchbase found that AI-driven startups raised $102 billion of that, or 64%, and that a record 27 companies raised rounds of $1 billion or more. Funding for the first nine months of 2026 came to $679 billion, the highest on record for that period.
Seed funding came to $13 billion in the quarter, according to Crunchbase, and about a fifth of it, $2.6 billion, went to rounds of $100 million or more. For a small team building trucking software or wiring an old SAP system into modern tools, free infrastructure, engineering help and an introduction to an investor can decide whether the company gets through its first year.
Corporates were involved in a record 82.6% of US venture deal value in the first half of 2026, according to PitchBook. Yet the Q2 2026 PitchBook-NVCA Venture Monitor found that corporate venture arms took part in only 21% of US deals, their lowest share in a decade, which PitchBook describes as "spending more on less". Affinity's 2026 corporate venture report found that 63% of corporate venture deals involve AI, compared with 49% for independent venture firms.
The market Confluent built is growing quickly
Mordor Intelligence estimates the streaming analytics market at $43.19 billion in 2026, rising to $175.55 billion by 2031, a compound annual growth rate of 32.38%. The firm notes that more than 80% of Fortune 100 companies rely on Kafka, while job boards list far more openings than qualified engineers. Confluent was founded in 2014 by Jay Kreps, Jun Rao and Neha Narkhede, the former LinkedIn engineers who built Kafka there as an open-source project.
Confluent's own 2026 Data Streaming Report, which surveyed 4,625 IT leaders, found that 88% rank data streaming as a high investment priority. It also found that 94% of technology leaders have seen or expect to see data streaming increase the impact of their AI investments, and that half of organisations report at least a fivefold return on their streaming spend. These are vendor-commissioned figures, but they explain why Confluent wants AI-native companies on its platform early.
"If you go look under the hood of all these frontier LLMs, in some shape or form they're all leveraging real-time streaming for fine-tuning, for feedback loops, for context engineering," said Kamal Brar, Senior Vice President of Global Partners and Technology Group at Confluent. "For us, we're in the thick of it."
The credits programme was a fix for a reputation
"Confluent had, or maybe still to a certain extent has, a reputation as a titanium-plated, enterprise-grade solution that isn't startup-appropriate. This is what Capital One uses, not a two-person startup," Graczewski said. "Part of the Confluent for Startups programme was to get over that perception."
The programme now gives up to $20,000 in Confluent Cloud credits for a year, plus time with a sales engineer and, where it makes sense, an account executive or local engineering support. In September 2023 Confluent said it had worked with more than 150 startups. By January 2026 it had received about 1,800 applications from 85 countries over three years and accepted roughly half. It allocates $500,000 to $600,000 a year in credits, and Graczewski said most of them are used.
"We don't accept everyone. We want to make sure you have a real use case and that it's viable," he said. "If you talk to other startup leaders, you give out credits and nobody ever uses them. We try to be very judicious with our credit allocation, and therefore they are used."
"I'll tell you it's an eight-figure return on a multi-hundred-thousand-dollar investment, in terms of total Confluent spend," Graczewski said. That is a return of at least $10 million on an annual credit budget below $1 million. The programme sits within corporate development, close to the top of the company.
"To an extent we've been very successful, but we still have a lot of work to go," he said of the effort to change how founders see Confluent.
The AI accelerator was built for speed
Confluent announced the Confluent for Startups AI Accelerator in October 2024 as a 10-week virtual programme for 10 to 15 early-stage startups building real-time, AI-powered applications on data streaming. Its benefits included early access to Confluent's newest AI and streaming technology, one-on-one technical and business mentoring with feedback on business models and pitch decks, marketing exposure, and credits from Confluent and partners including MongoDB and Anthropic. Applicants were judged on product vision, team and potential market impact.
"Early-stage startups are remarkable innovation labs, equipped with the focus and drive to build something disruptive without the burden of legacy," said Andrew Sellers, Head of Technology Strategy at Confluent, at the launch.
The first cohort, announced in February 2025, had 12 companies from five countries. They included Agent Taskflow and VytalSigns in New York, BioIntelliSense in Colorado, Lendica in Boston, Coxwave Align in Seoul, Flexprice and Nyaay AI in Delhi, Cowee, PricingOS and TwinLabs in London, and NomadicML and Reworkd in San Francisco. Their work ranged from AI agent management, health monitoring and billing infrastructure to dynamic ecommerce pricing and legal automation for India's courts.
"We joined Confluent's AI Accelerator because we see real-time data streaming as the key to truly continuous ML optimisation and personalisation," said Mustafa Bal, founder and CEO of NomadicML, at the time.
Cohorts of about 10 now run twice a year, from February to April and from September to November, which amounts to about 20 companies a year. That is a fraction of the roughly 900 startups accepted into the wider credits programme over three years. Each startup gets a dedicated technical mentor and works on an AI use case that depends on real-time streaming. Teams meet every two weeks, much of the work happens asynchronously on Slack, and each company builds a reference architecture with Confluent's engineers.
"These are our best people," Graczewski said. "The best people we have are working hand-in-hand with these startups."
Startups are Confluent's test bench
"Whereas the first model is a pipeline-generation, sales-led model, the accelerator is a product programme. It's about the lessons, the insights and the stories we can tell," Graczewski said.
In October 2025, at its Current conference in New Orleans, Confluent launched Confluent Intelligence, which included enhanced Streaming Agents and a Real-Time Context Engine. Streaming Agents run inside event streams so AI agents can monitor a business and act on events as they happen. The Context Engine, released in early access, is a fully managed service that streams materialised views to any AI agent or application, including through the Model Context Protocol.
"Our original hypothesis in the accelerator, which has been absolutely validated, is that small, smart teams who are very motivated to develop a great product quickly and get into production as fast as possible have had tremendous benefits for us," Graczewski said, adding that enterprise adoption of new features tends to be uneven. "In the last cohort, on our Slack channels, people were running into issues where we thought, we hadn't thought about that use case."
"A lot of the companies that we're talking about that are live in production are coming out of this programme," he said. "As startups, they get to market so much faster than a traditional enterprise, so we get tremendous value from the iterations we're able to get with them."
The accelerator recorded a Net Promoter Score of 100 across its first two cohorts. "Every startup that has come through has given us a nine or a ten on whether they would recommend this programme to another founder," Graczewski said. That score comes from about 20 companies, a small sample, and he said alumni posting on LinkedIn now bring in more applicants than any advertising does.
Selection rewards a defined problem
Prior experience with Kafka and Flink is "a huge leg up", Graczewski said, though not a requirement. What matters most is a specific use case, which Confluent tests through pre-screening and interviews.
"We can't take everyone," he said. Hemut fit because its product depends on time-sensitive events. "A load gets offered, who can pick this up? It's a perfect use case for us," Graczewski said. The company built its platform alongside the trucking firm it worked from, so the carrier could adopt it gradually. Confluent says Hemut once brought a carrier's full technology stack live in two days.
Anex, from the second cohort, takes legacy SAP data from sectors such as cement and moves it into a modern event-driven system. "It's not even tech 1.0," Graczewski said. "It's tech 0.1."
One first-cohort founder rebuilt his product on mentor advice
Saul Sparber spent about a decade as a data engineer, including consulting work for clients such as Nike, Pluto TV and Fox, before ChatGPT's release in December 2022 changed his plans.
"You would use ChatGPT in the web browser, it would spit out the response, and then you'd have to copy and paste it into whatever you were doing, your code editor, your email," Sparber said. "So I thought, why can't we just automate that process?"
He founded Agent Taskflow in 2023, building on Confluent from the start. "I didn't want to scale Kafka or do any of the DevOps," he said. During the first cohort, from January to April 2025, the company signed its first enterprise client, a $750 million healthcare company in New York.
"Full disclosure, my cohort was the first one, so they've had the opportunity to polish it up," Sparber said. "That being said, it was really good." The sessions covered fundraising, pitching and marketing, and a couple of sessions with Confluent's technical lead reshaped his architecture.
"Before, we had many topics, and the granularity was very small. We had different topics for events starting and completing and retrying," Sparber said. "We condensed it down into one topic per microservice. It cut down on our latency, it cut down on our build time, and it simplified the architecture, which made it a lot easier now that AI is doing a lot of the development."
"That pattern we worked through is still how we run our product today, over a year later," he added. Agent Taskflow then spent weeks retrofitting its product ahead of the demo and investor days.
The healthcare client pays for 25 seats at a negotiated rate of around $50,000 a year, and smaller accounts, including a solar marketplace and a telecom provider for multi-family buildings, pay about $10,000 a contract. Sparber has taken one $25,000 cheque and is otherwise bootstrapped. "We don't have huge revenue numbers right now, but we also haven't taken any huge funding," he said. "We are revenue positive."
Agent Taskflow's product has since moved towards governance, observability and running AI models locally on company devices, partly because of cost. "Their biggest pain point right now is, I'm spending too much on Anthropic tokens or on ChatGPT tokens," Sparber said of his clients. He plans to raise a formal round between late 2026 and the first quarter of 2027.
"There's this risk, every day, of waking up and one of the big giants like Anthropic or OpenAI has released a feature that is your product, and that has now happened thousands of times over to startups out there," he said.
A pitch day opens doors that cold emails do not
Each cohort ends with an internal demo day, followed two days later by a VC pitch day. Graczewski listed General Catalyst, Sequoia, Menlo Ventures and Bain Capital, alongside data-infrastructure specialists such as Tribe and Race Capital, or 15 to 20 firms in all.
"It's kind of an audition. After they pitch, hopefully the VCs will invite them back for a full half-hour pitch session," he said. "At least you get your shot of having them hear you."
He valued the mentoring alone at "tens or hundreds of thousands of dollars", given the seniority of the people involved.
Graczewski also invites niche investors such as Villain Capital, a vertical software fund named after the Dark Knight line "You either die a hero or see yourself live long enough to become a villain". It was founded by Tyler Sosin after 17 years at Menlo Ventures and Accel.
"His entire investment thesis is niche vertical SaaS operations, which other VCs are going to pass on because the TAM isn't big enough," Graczewski said. "It's probably going to become a cash cow. If you wrap it up and you own it, it becomes very, very interesting."
"If you get backed by Sequoia or Villain, it's all the same for this company in Mexico City that's doing charter buses for tourists," he added. "It all depends on your market."
Confluent already holds stakes in the winners
Confluent's first Data Streaming Startup Challenge, launched in September 2023, offered an investment of up to $500,000 for the winner and up to $250,000 each for two runners-up. It was open to companies under five years old that had raised less than $10 million, and the top 10 finalists each received $5,000 in credits. Confluent said the challenge drew nearly 100 applications from 22 countries.
London-based TwinLabs.ai, is a bootstrapped startup whose team learned to build real-time digital twins at the Tour de France, won the $500,000 seed investment in March 2024. TwinLabs later joined the accelerator's first cohort. The runners-up were Atomic Tessellator and Busie.
Busie, a Kingston, New York company founded in 2017 that makes charter bus management software, is the model Hemut followed. Graczewski said its founders embedded themselves in a small upstate bus operator, which is now where Busie has its headquarters.
Atomic Tessellator, which has moved from computational chemistry into materials discovery, raised $11.3 million in seed funding in April at a valuation of roughly $50 million to develop Vireon, a magnet that needs no rare earth elements. Crane Venture Partners led the round from its new $135 million Asia-Pacific fund, joined by In-Q-Tel, Icehouse Ventures, Outset Ventures and GD1, and Confluent returned as an existing backer.
"What we're showing is it is both," Graczewski said, referring to the view inside Confluent that the challenge was a marketing exercise. "Our stake in Atomic Tessellator is already going to have paid back the whole programme."
Focal Systems was Confluent's inaugural Americas Data Streaming Startup of the Year in 2025. Berlin-based Upvest, whose clients include DKB, N26 and Revolut, was named the first EMEA winner in July 2026. Graphiant, a San Jose network-as-a-service company founded in 2020, took the 2027 Americas title this month. InfiniteWatch, a New York customer-intelligence startup, joined Confluent for Startups in the fourth quarter of 2025 and had its product in production by the end of the first quarter of 2026.
Rivals have built similar paths from credits to capital
Graczewski traced the model back about a decade to AWS, where he once covered Y Combinator, and credited Snowflake with leading the data-infrastructure companies within the past five years. "Now it's becoming credits plus," he said. "The plus is curated."
More than 330,000 startups have joined AWS Activate, according to AWS. It offers $1,000 to self-funded founders, up to $5,000 for some, up to $200,000 for startups backed by an approved provider, and more for AI startups ready to scale. In 2024 AWS committed $230 million to generative AI startups. Its eight-week Generative AI Accelerator gives each selected company up to $1 million in credits, and its 2025 cohort had 40 startups.
Google offers VC-backed, AI-first startups up to $350,000 in credits over two years, covering 100% of eligible usage in the first year up to $250,000. Nvidia's Inception programme is a membership network rather than a selective cohort, and Nvidia says it has more than 40,000 members.
In July 2025 Microsoft replaced a model that had offered AI-focused companies up to $150,000 in Azure credits. Teams with a linked investor now start at $100,000, and unfunded teams can get up to $5,000.
Databricks relaunched its startup programme in June 2026 with up to $200,000 in credits across Databricks and Neon, its Postgres database. In September 2025 it launched an invitation-only AI Accelerator for pre-seed and seed companies, powered by Databricks Ventures, that offers up to $250,000 in funding and product credits. Databricks Ventures, launched about four years ago, now has more than 50 portfolio companies and makes roughly 15 new investments a year.
"We're more focused on whether it makes strategic sense to get into the round, than on the valuation," said Andrew Ferguson, who leads Databricks Ventures. "We don't lead rounds."
Snowflake combined its efforts as Snowflake for Startups in September 2025. It said its accelerator had helped participants launch 28 live products, with applications up 304%, and that Snowflake Ventures planned to increase its investment pace by more than 30% and grow its portfolio and exits to more than 65 companies. Its annual Startup Challenge, run with the New York Stock Exchange, offers finalists a share of up to $1 million from Snowflake Ventures. The 2026 finalists were Airrived, LGND AI and Twine Security.
MongoDB added Temporal and Fireworks AI as partners offering matched credits in January, and said companies in MongoDB for Startups represent more than $200 billion in total valuation, based on PitchBook data. "They need a robust data foundation and a stack that works from day one and scales with their business," said Suraj Patel, VP, MongoDB Ventures & Corporate Development.
Confluent's $20,000 credit cap is a fraction of what hyperscalers offer. It offers no cash at the accelerator stage, while Databricks does, and its pitch day is an introduction rather than an investment. Graczewski's argument is that focus is the advantage: credits that actually get used, mentors drawn from Confluent's most senior engineers, and a programme built around one demanding type of workload.
"MongoDB used to give away whatever it was, like $2,000 in credits, and all you needed was a website. They've since clawed that back and are making it harder to get in, but you're getting more resources," Graczewski said. "These startup programmes aren't for everyone."
Free compute comes with a cliff
"Most startups burn through compute credits not because they're building too much," said Ashutosh Prakash Singh, CEO of Revrag, adding that "without discipline it quickly turns into uncontrolled spend."
"Working without cloud credits is another one of those choices that make our business model and cost of goods sold real," said Praveer Kochhar of KOGO AI, whose company has chosen to build without them.
"Don't aim to spend $20,000 in credits, because that means you're probably going to spend $20,000 with us the next year, and you probably don't want to do that if you can help it," Graczewski said of the advice he gives his own founders.
"Yes, we want you to use Confluent and spend with us, but we want it to be as efficient and effective as possible, because data streaming drags forward compute and storage," he added, explaining why the accelerator builds reference architectures with each startup. "The credits are going to run out."
"If you don't have a high-volume, low-latency, real-time data requirement, don't use us," Graczewski said. "If you're going to be streaming telemetry from a thousand Nvidia edge servers and you need to make decisions in milliseconds, then yes, we need to be part of what you're doing."
The technical founder is the real audience
Graczewski estimated that a seed-stage founder could combine around a quarter of a million dollars in AWS credits, $20,000 from Confluent, $50,000 from Databricks and $15,000 to $20,000 from MongoDB. "When I came to Silicon Valley, if you wanted to do something, you had to buy Sun servers. It was an enormous capital expense," he said. "Not only has cloud infrastructure reduced that, but the startup programmes I'm part of have taken it to zero for the first year of development."
"Who's your technical founder, who's going to show you how you're going to do it, and that this is going to scale effortlessly?" Graczewski said, describing what investors now ask. Hemut had 200 trucks on its system and was planning for thousands. "When that happens, is your platform going to be ready, or did you jury-rig this thing so that once it becomes complex the whole thing crashes?"
"You go to TechCrunch Disrupt, Startup Grind, Techstars, everything is geared towards the founder CEO who is the evangelist, who's raising the money," he said. "There's almost nothing, as far as I can tell, as a platform for the technical founder who is building the product."
He is working with his counterparts at Databricks, MongoDB and Datadog on what he calls a "builder path", even though their companies compete. "There are some areas where a data warehouse is better than a data lake," he said. "There's a partner for you out there, and it's probably a package of partners."
"From day zero, it needs to be something you're thinking about thoughtfully," he added. "It needs to be as thoughtful as your product-market fit, as your go-to-market."
IBM's ownership adds a new layer
IBM announced on December 8, 2025 that it would buy Confluent for $31 a share in cash, valuing the company at about $11 billion, a premium of nearly 34% over Confluent's previous closing price. IBM completed the purchase on March 17, 2026, and said Confluent had more than 6,500 enterprise customers, including 40% of the Fortune 500.
"IBM and Confluent together will enable enterprises to deploy generative and agentic AI better and faster by providing trusted communication and data flow," said Arvind Krishna, IBM's chief executive. IBM said Confluent's addressable market had doubled to $100 billion since 2021, and listed day-one integrations with watsonx.data, IBM MQ, webMethods Hybrid Integration and IBM Z.
"[Confluent] fills a critical gap within the watsonx platform, IBM's next-gen AI platform, by providing the ability to monitor real-time data," said Scott Bickley, advisory fellow at Info-Tech Research Group, who also noted that Confluent had been experiencing slowing revenue growth. Forrester's Mike Gualtieri and David Mooter noted that IBM has not committed to keeping Confluent as a distinct business, and expect the roadmap to tilt towards IBM's own products.
The Confluent for Startups team was not authorised to discuss the acquisition for this story, and the interviews were conducted between the deal's announcement and its completion. Since March, Hemut's cohort has finished on an IBM-owned platform, Confluent has returned as a backer in Atomic Tessellator's April round, and Confluent has named startup award winners in July and this month.
The next phase is venture capital
Graczewski expects startup programmes to keep moving from marketing to product, and then to investment. As of January, Confluent had no formal venture arm, while MongoDB, Databricks and Snowflake did, though its challenge stakes show it already invests.
"Companies are going to realise they have an unfair advantage in investing because of the insight they have into infrastructure and consumption," he said. "I think you're going to see data infrastructure companies develop very meaningful venture programmes, $250 million funds putting in $5 million cheques, and being very successful at it."
"The fact that we now have over 50 portfolio companies actually compounds that learning benefit," Databricks' Ferguson has said of the same advantage. Confluent's new owner already runs a $500 million Enterprise AI Venture Fund that writes cheques of $2 million to $5 million, though IBM has not said whether it will be connected to Confluent's startup programme.
A vendor that mentors a startup, sees its usage data and may later own part of it ends up as supplier, adviser and potential shareholder at once. Hemut, Agent Taskflow and Atomic Tessellator show what Confluent's model can offer. Founders deciding how far to tie their architecture, roadmap and cap table to one platform will have to weigh that offer against what they give up.