Deepgram.

The company you are already talking to.

A drive-thru in Ohio is taking your order. The voice is calm, a little upbeat, faintly recognizable as not-quite-human. It hears you mumble. It catches you mid-sentence when you change your mind from a milkshake to a malt. It does not, crucially, sound like a robot reading a script in 1998.

Somewhere in a data center, Deepgram is doing the heavy lifting. Speech in. Meaning out. Speech back. Repeat, 80 million times a day, across customer service, healthcare intake, meeting notes, dispatch lines, and a creeping number of consumer apps that have quietly grown ears.

The company is now eleven years old, a fresh unicorn at $1.3 billion, and increasingly the answer to a question a lot of CTOs are now asking out loud: who is going to power the voice layer of AI? Deepgram would like that to be Deepgram.

The problem they saw.

For most of the 2010s, speech recognition was a feature you tolerated rather than a product you loved. Cloud providers bolted aging acoustic models onto cleaner APIs and called it modern. Accuracy was decent for American English in a quiet room. Anywhere else - a call center floor, an accent the training set didn't love, three people talking over each other in a Zoom - it broke politely.

The market had agreed, quietly, that this was fine. Deepgram disagreed, loudly.

The pitch was almost stubbornly simple: train an end-to-end deep learning system on enormous quantities of audio, ship it as an API, and beat the incumbents on accuracy, latency, and price. Not one of those three. All three.

The founders' bet.

Three physicists walk into a startup. It is not the setup of a joke, although it should be.

Scott Stephenson, Adam Sypniewski, and Noah Shutty met at the University of Michigan, where Stephenson did a PhD in particle physics. His thesis work involved building a lab two miles underground - the kind of place where you go to detect particles that have, statistically, almost no interest in being detected. Dark matter. The universe's most reluctant subject.

Audio, by comparison, is generous. It is everywhere. It is messy. It is exactly the kind of signal problem a former dark matter physicist looks at and thinks, this is downhill.

They went through Y Combinator in 2016 with a thesis the market thought was already settled - and the next ten years of accuracy benchmarks suggested otherwise. YC W16 end-to-end

What they ship.

Deepgram's product surface is, on paper, three letters plus a couple of greek-looking names. In practice it is the plumbing for any company that wants to put a microphone in front of an AI and get a useful answer back.

What is unusual about all of this, in 2026, is that Deepgram trains its models end to end rather than fine-tuning somebody else's open checkpoints. That choice was contrarian in 2015 and is increasingly contrarian again now that everyone is wrapping somebody else's API. It is also why the company can charge a fraction of incumbent prices and still operate a real business.

The proof.

It is one thing to claim better accuracy. It is another to win the kinds of customers who measure that claim with stopwatches and statisticians.

The 1,300-customer logo wall does the heavy lifting in pitch decks. The Series C, though, did something more telling: it brought Twilio, ServiceNow, SAP, and Citi Ventures onto the cap table as strategic investors. That is what it looks like when the companies that would normally build voice AI in-house decide it is cheaper, faster, and more accurate to just route the audio to Austin.

Why they say they are doing this.

Mission statements in voice AI tend toward the grand and the unfalsifiable. Deepgram's is mercifully concrete: make every voice heard and understood by machines. Translation: a janitor in Manila and a cardiologist in Boston should get the same accuracy on the same hardware at the same price.

That phrasing is not entirely altruistic. The companies that can transcribe and respond to every voice - not just the ones that show up in the training data - capture the rest of the market. Doing the right thing happens to be the most defensible thing. The Venn diagram is two circles.

The culture

The company runs lean for its scale. Roughly 210 people, hubs in San Francisco and Austin, a remote-friendly engineering culture descended from research labs. Internal tooling skews toward research dashboards rather than product analytics. Notion, Linear, GitHub, Slack. The model training infrastructure is genuinely impressive and almost nobody outside the company gets to see it.

Why it matters tomorrow.

Here is the bet, in case you would like to take the other side of it.

Every consumer app eventually becomes a conversation. Every enterprise workflow that touches a human eventually has a voice agent in it. Every meeting becomes a searchable transcript. Every call center hires fewer humans and more APIs. The platform that owns the underlying voice models gets to be a kind of toll booth on that traffic.

OpenAI, Google, Microsoft, ElevenLabs, AssemblyAI - the field is crowded and well-funded. Deepgram's argument is the only one a serious infrastructure company can make: we own the models, we serve them efficiently, we are not anyone's feature.

The drive-thru, revisited.

Back to the drive-thru. The line moves faster than it used to. The order takes fewer corrections. The person inside the speaker - if that is still the right word - hears your accent without flinching and politely waits while a toddler in the backseat changes her mind about cheese.

Eleven years ago, this conversation did not happen. Not because nobody wanted it to. Because the technology to do it well did not exist, and the incumbents had quietly agreed not to build it.

Three physicists thought that was a strange thing to agree about. They built it instead.

— END OF FILE —

Find them.

Watch & listen