By Ahmad IqbalAbout 18 minutes to read

The thesis

Investment Thesis

Exploratory project. Not a fund. Not raising.

BlackBerry couldn't see past the mobile keyboard.

What happens if Uber, YouTube, Duolingo, Figma or Apple can't see past the screen?

The claim

What if the keyboard was the detour?

Within a decade, speech will take back a large share of how people create and communicate with technology, after five centuries in which the printed and written word held almost all of it.

This is a retrieval (not a replacement). Writing does not go away, any more than typing went away when the camera took over sharing. The spoken word simply wins back ground it lost to print, and the medium becomes the message again: when the interface is a conversation, everything built on it starts to behave like one. We will know we are right when the keyboard and mouse feel like the landline: still in the home, but rarely the first thing anyone reaches for.

Three ages of the word

Primary orality

For most of human history, knowledge had a pulse. It lived in people. Every story, law and skill was spoken, remembered and passed on in conversation, and when the person died, the knowing died with them unless someone had listened. Walter Ong called this primary orality. The village organised itself around the elder, because the elder was the library. You did not look things up. You went and asked, and the answer came with a story attached.

Secondary orality

Writing broke the spell. Once a law could be written down, a person with no mistakes could become a judge: they did not need the elder's life, only the elder's text. Knowledge moved out of people and onto the page, and society reorganised around the page. Books, codes, encyclopaedias, contracts. Writing changed not just how we store thought but how we think, bringing distance, analysis and the fixed text. Five centuries of print made the eye the boss of the other senses.

Then radio and television brought speech back, but speech built on top of writing, and running one way. Ong called this secondary orality. You could listen, but you could not answer.

Hybrid orality

We are now entering a third age. You talk, and it shows. Technology can hear what you say, understand it, answer back and act on it. The back-and-forth that broadcast removed has returned, and this time the machine takes part in the conversation. McLuhan's word for this is retrieval: an old form comes back, changed, inside a new medium. The elder returns, but now the elder is everywhere, and everyone can ask. We call it Hybrid orality.

The mandate

The Spoken Fund backs the tools and companies of Hybrid orality: products where speech and conversation are the primary way people create, learn, decide and act, and where the technology answers back. Not voice bolted onto a screen. Voice as yang.

Stage

Early

Region

Global, with a lean towards the Gulf and Asia: markets where voice notes dominate, and where keyboards built for English fit the local language poorly (Arabic, for example).

Cheque sizes

US$250,000 to $1,000,000

Retrievals

What do the emoji and the podcast have in common?

This is the rule the whole thesis rests on. When a new medium wins, the old one does not die. It goes quiet, then comes back through the new medium, bigger than it ever was. Every new medium arrives wearing the old one's clothes. Then it finds its own.

The emoji is a hieroglyph.

For five thousand years we moved from pictures to letters, letter by letter. Then the digital keyboard handed us a thumbs up, a heart, a face in tears, and we took them gladly. One symbol now does the work of a sentence, and often does it better. The oldest writing system on earth came back as the newest, from a generation thousands of years removed from the age of hieroglyphs, and nobody needed to vote on it.

Radio read the newspaper aloud.

The first radio hosts picked up the morning paper and read it, top to bottom. Secondary orality was born as a man reading print into a microphone. Then radio discovered what the newspaper could never do. It could play music. It could take a call. The listener phoned in, and the audience became part of the show. The consumer edged towards the producer's chair.

Television took the microphone back.

Scripted, recorded, one way. Reality TV was the exception that showed the appetite. Then livestreaming arrived, with the chat scrolling beside the host, and the call-in show returned at the scale of tens of millions. Participation with content keeps coming back because people keep wanting it. Some of the most popular YouTube shows are a retrieval of the radio talk show, which itself was a retrieval of engaging with your family over the dinner table/campfire.

The pattern is McLuhan's: the first content of a new medium is always the old medium, until the new one finds its native form. Most AI voice today, from speech-to-text to text-to-speech, is at the newspaper-on-the-radio stage. It still orbits the written word.

We are looking for its native forms: voice applications that turn people back towards the physical world and the people in it.

Why it matters

Why is the most connected generation afraid of a phone call?

Text has overheated.

A generation raised on messaging can take its time, edit before sending, or leave a message on read, none of which a face-to-face conversation allows. Many now find a live call hard. In one UK survey, nearly 70% of 18 to 34 year olds preferred a text to a call (source (opens in a new tab)), and nearly a quarter never answer their phone (source (opens in a new tab)). In another, 58% of young workers script or rehearse a conversation before a call (source (opens in a new tab)). In South Korea, the share of young people reporting call phobia rose from 30% in 2022 to 41% in 2024 (source (opens in a new tab)).

McLuhan's rule is that a medium pushed to its extreme flips into its opposite: the car promised freedom until every road filled with traffic, and the car became the thing that stops you moving. Text was pushed that far. Every relationship, all day, in writing. And it flipped: the voice call, once the ordinary way to reach someone, became a source of anxiety.

The flip back has begun. The same generation sends more voice notes than any other: 85% of Gen Z leave them, against 45% of Baby Boomers (source (opens in a new tab)), and 37% of 18 to 34 year olds prefer a voice note to a call (source (opens in a new tab)). They have not given up on the voice. They want it back on their own terms.

The sceptic says Gen Z chose text because it is better. We believe they chose it because it was the only channel that gave them control. A call means performing live. Text gives you time. People were not fleeing speech, they were fleeing the live performance. Voice AI removes that pressure. No one judging, no wrong time, no awkward silence. Speech with the control of text. The voice note was the first sign. Voice agents are the release valve.

A second reversal is under way, in how business reaches people. Business-to-consumer assumes the journey starts with the business, and the work is building a path to the customer: the advert, the app, the checkout. When a person's own agent does the shopping, the journey starts with the consumer. The business stops building paths to people and starts making itself easy for their agents to find, trust and buy from. The first signs are here: in September 2025 OpenAI and Stripe released an open standard that lets an AI agent complete a purchase with a shop inside a conversation (source (opens in a new tab)).

Speaking is how most people think out loud, build relationships and trust, teach, sell and decide. Tools built around conversation turn that into finished work, and people already pay for them. That is where a first fund on this thesis would put most of its money.

There is a longer-term prize as well. The spoken word is more accessible than the written word, and hundreds of millions of people cannot read or write well enough to take part in a digital economy built on text. As voice gets cheaper, speech can bring them in. We treat that as the next phase of this market... maybe a fund 2 focus.

Why now

AI voice is getting cheap quickly. Listening has been close to free since 2023 (source (opens in a new tab)). The price of a minute of AI speech has fallen by 60% or more at OpenAI (source (opens in a new tab)) and ElevenLabs (source (opens in a new tab)) since late 2024, and newer tools such as Deepgram (source (opens in a new tab)), Cartesia (source (opens in a new tab)) and Google's Gemini Live (source (opens in a new tab)) have launched at lower prices.

AI voice prices fall fast early, and newer tools launch cheaper

List price per minute of AI voice, US cents. Phone network costs excluded.

  • OpenAI Realtime
  • ElevenLabs
  • Newer tools: Deepgram, Cartesia, Gemini Live
  • OpenAI Whisper (listening)
0¢5¢10¢15¢20¢25¢2023202420252026OpenAI Whisper (listening): 0.6¢ a minute, Mar 2023OpenAI Whisper (listening): 0.3¢ a minute, Mar 2025Gemini Live (speaking): 1.8¢ a minute, Sep 2025Cartesia agent: 6¢ a minute, Jul 2026Cartesia agent: 3¢ a minute, Oct 2026Deepgram agent: 7.5¢ a minute, Jun 2025ElevenLabs agent: about 20¢ a minute, Dec 2024 (implied, not published)ElevenLabs agent: 10¢ a minute, Feb 2025ElevenLabs agent: 8¢ a minute, May 2026OpenAI Realtime (speaking): 24¢ a minute, Oct 2024OpenAI Realtime (speaking): 9.6¢ a minute, Dec 2024OpenAI Realtime (speaking): 7.68¢ a minute, Aug 202524¢~20¢ (implied)0.6¢ElevenLabs 8¢OpenAI Realtime 7.68¢Deepgram 7.5¢Cartesia 3¢Gemini Live 1.8¢Whisper (listening) 0.3¢0¢5¢10¢15¢20¢25¢2023202420252026OpenAI Whisper (listening): 0.6¢ a minute, Mar 2023OpenAI Whisper (listening): 0.3¢ a minute, Mar 2025Gemini Live (speaking): 1.8¢ a minute, Sep 2025Cartesia agent: 6¢ a minute, Jul 2026Cartesia agent: 3¢ a minute, Oct 2026Deepgram agent: 7.5¢ a minute, Jun 2025ElevenLabs agent: about 20¢ a minute, Dec 2024 (implied, not published)ElevenLabs agent: 10¢ a minute, Feb 2025ElevenLabs agent: 8¢ a minute, May 2026OpenAI Realtime (speaking): 24¢ a minute, Oct 2024OpenAI Realtime (speaking): 9.6¢ a minute, Dec 2024OpenAI Realtime (speaking): 7.68¢ a minute, Aug 202524¢~20¢ElevenLabs 8¢OpenAI 7.68¢Deepgram 7.5¢Cartesia 3¢Gemini 1.8¢Whisper 0.3¢

Hollow dot: ElevenLabs launch price, implied rather than published. Lines run flat to today where no newer price was found.

Show the numbers
AI voice prices
SeriesDatePrice per minute
OpenAI Realtime (speaking)Oct 202424¢
OpenAI Realtime (speaking)Dec 20249.6¢
OpenAI Realtime (speaking)Aug 20257.68¢
ElevenLabs agentDec 2024~20¢ (implied)
ElevenLabs agentFeb 202510¢
ElevenLabs agentMay 20268¢
Deepgram agentJun 20257.5¢
Cartesia agentJul 20266¢
Cartesia agentOct 20263¢
Gemini Live (speaking)Sep 20251.8¢
OpenAI Whisper (listening)Mar 20230.6¢
OpenAI Whisper (listening)Mar 20250.3¢
List prices from vendor announcements and pricing pages, March 2023 to November 2026 · OpenAI and Gemini per-minute figures converted from token prices · Cartesia, Deepgram and Gemini figures still to be checked against each vendor's own page

Two years after launch, text had lost over 99% of its price. Voice has lost 60 to 70%.

Left axis: GPT-4-quality text, $ per million tokens. Right axis: AI voice, cents per minute.

  • GPT-4-quality text (left axis)
  • OpenAI Realtime (right axis)
  • ElevenLabs (right axis)
$0$10$20$30$400¢5¢10¢15¢20¢25¢Launch1 year2 years3 yearsYears since launch← Text, $ per million tokensVoice, ¢ per minute →ElevenLabs agent: about 20¢ a minute, Dec 2024 (implied)ElevenLabs agent: 10¢ a minute, Feb 2025ElevenLabs agent: 8¢ a minute, May 2026OpenAI Realtime (speaking): 24¢ a minute, Oct 2024OpenAI Realtime (speaking): 9.6¢ a minute, Dec 2024OpenAI Realtime (speaking): 7.68¢ a minute, Aug 2025GPT-4-quality text (GPT-4): $37.50 per million tokens, Mar 2023GPT-4-quality text (GPT-4 Turbo): $15.00 per million tokens, Nov 2023GPT-4-quality text (GPT-4o): $7.50 per million tokens, May 2024GPT-4-quality text (GPT-4o mini): $0.26 per million tokens, Jul 2024GPT-4-quality text (Gemini 2.0 Flash): $0.18 per million tokens, Feb 2025GPT-4-quality text (GPT-5 nano): $0.14 per million tokens, Aug 2025GPT-4-quality text (Qwen3.7 Flash): $0.055 per million tokens, Jul 2026Launch: $37.50 text, 24¢ voiceOpenAI Realtime · 7.68¢ElevenLabs · 8¢Text · $0.055$0$10$20$30$400¢5¢10¢15¢20¢25¢Launch1 yr2 yr3 yrYears since launchElevenLabs agent: about 20¢ a minute, Dec 2024 (implied)ElevenLabs agent: 10¢ a minute, Feb 2025ElevenLabs agent: 8¢ a minute, May 2026OpenAI Realtime (speaking): 24¢ a minute, Oct 2024OpenAI Realtime (speaking): 9.6¢ a minute, Dec 2024OpenAI Realtime (speaking): 7.68¢ a minute, Aug 2025GPT-4-quality text (GPT-4): $37.50 per million tokens, Mar 2023GPT-4-quality text (GPT-4 Turbo): $15.00 per million tokens, Nov 2023GPT-4-quality text (GPT-4o): $7.50 per million tokens, May 2024GPT-4-quality text (GPT-4o mini): $0.26 per million tokens, Jul 2024GPT-4-quality text (Gemini 2.0 Flash): $0.18 per million tokens, Feb 2025GPT-4-quality text (GPT-5 nano): $0.14 per million tokens, Aug 2025GPT-4-quality text (Qwen3.7 Flash): $0.055 per million tokens, Jul 2026$37.50 · 24¢OpenAIElevenLabsText

Different units on one picture: compare the shapes of the lines, not their heights.

Show the numbers
Text and voice prices since launch
SeriesDatePrice
GPT-4-quality textMar 2023$37.50 per million tokens
GPT-4-quality textNov 2023$15.00 per million tokens
GPT-4-quality textMay 2024$7.50 per million tokens
GPT-4-quality textJul 2024$0.26 per million tokens
GPT-4-quality textFeb 2025$0.18 per million tokens
GPT-4-quality textAug 2025$0.14 per million tokens
GPT-4-quality textJul 2026$0.055 per million tokens
OpenAI RealtimeOct 202424¢ per minute
OpenAI RealtimeDec 20249.6¢ per minute
OpenAI RealtimeAug 20257.68¢ per minute
ElevenLabsDec 2024~20¢ per minute (implied)
ElevenLabsFeb 202510¢ per minute
ElevenLabsMay 20268¢ per minute
Text from The Brief Script AI API price index, launch March 2023 · OpenAI Realtime launch October 2024 · ElevenLabs agents launch December 2024, launch price implied · voice has about two years of data

The chart puts two different units on one picture, so compare the shapes of the lines, not their heights.

Text shows how fast this technology can get cheaper. Voice is on the same underlying technology but is falling more slowly so far, and the gap between the two is the room voice has to fall. Whether it closes is the bet.

Read moreShow less

The two speaking lines are not like for like. OpenAI's price covers the model that thinks and speaks, while ElevenLabs' covers the voice layer and bills the language model on top (source (opens in a new tab)). At 8 cents a minute (source (opens in a new tab)), an hour of conversation still costs about $4.80, so the access cases in this thesis depend on prices continuing to fall.

Prices have fallen in steps, not in a smooth line. Each tool was cut sharply in its first months and has moved little since.

We found no study that forecasts the price of AI voice over the next decade. The closest evidence is for AI text. Andreessen Horowitz estimates that the price of a text model of equal quality falls about tenfold a year (source (opens in a new tab)). Stanford's AI Index measured a 280-fold fall in under two years, from $20 to 7 cents per million tokens between November 2022 and October 2024 (source (opens in a new tab)). Voice models are built on the same technology, so we expect voice to follow. That is our inference, not a published forecast.

The price of GPT-4 quality text has fallen about 680-fold since 2023

Cheapest model at GPT-4 quality or better. US dollars per million tokens, blended.

$0$10$20$30$402023202420252026GPT-4, Mar 2023: $37.50 per million tokensGPT-4 Turbo, Nov 2023: $15.00 per million tokensGPT-4o, May 2024: $7.50 per million tokensGPT-4o mini, Jul 2024: $0.26 per million tokensGemini 2.0 Flash, Feb 2025: $0.18 per million tokensGPT-5 nano, Aug 2025: $0.14 per million tokensQwen3.7 Flash, Jul 2026: $0.055 per million tokensGPT-4 $37.50 · Mar 2023Qwen3.7 Flash $0.055 · Jul 2026$0$10$20$30$402023202420252026GPT-4, Mar 2023: $37.50 per million tokensGPT-4 Turbo, Nov 2023: $15.00 per million tokensGPT-4o, May 2024: $7.50 per million tokensGPT-4o mini, Jul 2024: $0.26 per million tokensGemini 2.0 Flash, Feb 2025: $0.18 per million tokensGPT-5 nano, Aug 2025: $0.14 per million tokensQwen3.7 Flash, Jul 2026: $0.055 per million tokensGPT-4 $37.50Qwen3.7 Flash $0.055

In between: GPT-4 Turbo $15 (Nov 2023), GPT-4o $7.50 (May 2024), GPT-4o mini $0.26 (Jul 2024), Gemini 2.0 Flash $0.18 (Feb 2025), GPT-5 nano $0.14 (Aug 2025).

Show the numbers
Price of GPT-4-quality text
DateModelUS$ per million tokens
Mar 2023GPT-4$37.50
Nov 2023GPT-4 Turbo$15.00
May 2024GPT-4o$7.50
Jul 2024GPT-4o mini$0.26
Feb 2025Gemini 2.0 Flash$0.18
Aug 2025GPT-5 nano$0.14
Jul 2026Qwen3.7 Flash$0.055
The Brief Script AI API price index · blended price is three parts input to one part output · GPT-4 quality is a score of 126 or more on the Epoch Capabilities Index

This is the price of a fixed level of quality, not of the newest model. OpenAI's newest flagship models cost more than they did a year ago (source (opens in a new tab)). What collapses is the price of yesterday's best.

Voice is not at parity with text. On OpenAI's current real-time model, a minute of the AI speaking costs about 7.7 cents (source (opens in a new tab)). The same words as text cost about half a cent, roughly 16 times less. Audio is priced higher per unit, and speech needs about six times as many units to say the same thing. The 16-times figure is our estimate, at 150 words and about 200 text tokens a minute.

The Jevons case

Cheaper voice should grow the market for voice, not shrink it. Coal taught us this once already.

In 1865 the economist William Stanley Jevons observed that as steam engines burned coal more efficiently, Britain used more coal, not less (source (opens in a new tab)). Each gain in efficiency made coal pay in uses where it had not paid before. This is known as the Jevons paradox (source (opens in a new tab)).

Engines needed far less coal for the same work, and Britain burned far more of it

Bars: UK coal output, million tonnes (left). Line: pounds of coal per horsepower-hour (right).

  • UK coal output (bars, left axis)
  • Coal per horsepower-hour (line, right axis)
01002003000 lb4 lb8 lb12 lb180018501900← Coal output, million tonnesPounds of coal per horsepower-hour →UK coal output, 1800: 10 million tonnes10 MtUK coal output, 1853: 71 million tonnes71 MtUK coal output, 1913: 292 million tonnes292 Mt1800: 12 lb of coal per horsepower-hour1865: 4 lb of coal per horsepower-hour1890: 1.5 lb of coal per horsepower-hour1911: 1.31 lb of coal per horsepower-hour1920: 1.2 lb of coal per horsepower-hour12 lb4 lb1.5 lb1.2 lb010020030004812180018501900UK coal output, 1800: 10 million tonnes10UK coal output, 1853: 71 million tonnes71UK coal output, 1913: 292 million tonnes2921800: 12 lb of coal per horsepower-hour1865: 4 lb of coal per horsepower-hour1890: 1.5 lb of coal per horsepower-hour1911: 1.31 lb of coal per horsepower-hour1920: 1.2 lb of coal per horsepower-hour1241.51.2

Coal output bars for 1800, 1853 and 1913. Engine figures for 1800, 1865, 1890, 1911 and 1920.

Show the numbers
UK coal output
YearMillion tonnes
180010
185371
1913292
Coal per horsepower-hour
YearPounds of coal per horsepower-hour
180012
18654
18901.5
19111.31
19201.2
Coal output: 1913 peak from Our World in Data, 1800 and 1853 as quoted in Rebecca Solnit's Orwell's Roses · Coal per horsepower-hour: Clark and Jacks (1800), Jevons (1865), Technics History (1890), Encyclopaedia Britannica 1911, NavWeaps (1920) · all still to be checked against primary series

Chart sources: (source (opens in a new tab)) (source (opens in a new tab)) (source (opens in a new tab)) (source (opens in a new tab)) (source (opens in a new tab)) (source (opens in a new tab))

We expect voice to follow the same pattern. At 24 cents a minute, few uses could justify AI voice. Each price cut brings in uses that could not turn a profit before, and the access cases in this thesis sit at the far end of that curve.

This is a bet, not a law. The paradox holds only when demand responds strongly to price. The evidence that would confirm it is total minutes of AI voice growing faster than the price per minute falls.

Early evidence points the same way. ElevenLabs cut its agent price from about 20 cents a minute at launch in late 2024 to 8 cents in 2026. Over roughly the same period its annual recurring revenue rose from about $200 million in September 2025 (source (opens in a new tab)) to about $600 million in September 2026 (source (opens in a new tab)), and its agents now handle more than 15 million conversations a week (source (opens in a new tab)).

(If anyone reading this can get me on the ElevenLabs cap-table, please reach out, I'm buying)

The long-term opportunity: access

Every child learns to talk. Typing takes a school.

Nobody is taught to speak. Everybody is taught to read and write, and not everybody gets the chance. Voice is the only interface with no onboarding, which is why the people furthest from the keyboard are the ones it reaches first.

This is the long-term case, and the clearest example of the Jevons effect. It is not where a first fund would put most of its money.

The biggest group of people left out of the digital economy already lives within range of a mobile network. What they lack is the ability and confidence to use it.

3.1 billion people have mobile internet coverage but are not online

Nearly 4 in 10 people in the world

GSMA, 2026 · 3.1 billion people, 38% of the world's population

The network is not what stops them. The two biggest barriers are the cost of a phone and a lack of literacy and digital skills (source (opens in a new tab)). The internet they are offered has to be read and typed, and 739 million adults cannot read or write a simple sentence (source (opens in a new tab)).

Nobody has published how many of the 3.1 billion are held back by literacy alone, so we do not claim a number. GSMA estimates that bringing this group online would add about $3.5 trillion to the world economy between 2023 and 2030 (source (opens in a new tab)). Speech removes the reading barrier, and that is the part of the gap this thesis is about.

Who pays

The thesis stands on the two pillars of any economy: creation and consumption. Both pay now, and both are where a first fund would invest. The Access play comes later.

Creation.

People who can read and type, have a want to create and be creative, but don't have the technical skills to create or would rather talk. For example you want to publish a new offer for your voice agent to sell, instead of trying to visually communicate this, you simply leave a voice note for your agent. Speaking makes creating easier and more natural, and what comes out sounds like the person who made it. They already own phones, have audiences and pay for creative tools today.

Consumption.

People who would rather ask for what they want than work an app. Voice does not interrupt. You tell one assistant, your chief of staff, that the house is out of milk and what you want for dinner on Friday, and you carry on with your day. It passes that to a grocery agent, which places the order directly with the shop's own systems. The consumer talks to an agent, and the agent deals with the business. These customers already pay for the groceries, the rides and the subscriptions, so the money is there today. What changes is who owns the conversation.

Both Creation and Consumption use cases come down to free hands. One theory of how language began is that speech freed our hands from gesturing, so people could talk and work at once. Voice does it again. When neither making nor buying needs your eyes or your hands, people can do more of both in the same day, and sometimes in the same moment: a conversation that is also the finished work is creation and consumption happening at once. We believe this will have profound cultural impact to society.

Access.

People the written internet left out. Purchasing power is lower in these markets, but high-value uses exist there now. A farmer who needs crop market prices and advice will pay a rate at which that information earns more than it costs. As the price of a minute of voice keeps falling, more of these uses become worth paying for, which is the Jevons case above.

The evidence so far is mixed, and it supports the Jevons point. A two-year trial of a voice advice service for cotton farmers in India found strong demand and trust, but farmers were willing to pay less than the service then cost to run (source (opens in a new tab)). What has changed is the cost. Digital Green reports that its AI assistant for farmers now answers a question for under four cents (source (opens in a new tab)) and supports a farmer for a year for under a dollar (source (opens in a new tab)).

Where the value sits

Why don't humans have earlids, when we do have eyelids?

Because hearing is the sense that kept us alive while our eyes were busy or shut. Sound is the always-on channel, by design. Every screen app is fighting for the same pair of eyes, and the fight is the business model. A voice doesn't need your eyes. It can be with you while you cook, drive or walk, without taking anything away. That is a different kind of relationship, and it is the moat we care about most.

We back the products built on voice, not the companies selling the models. The charts above show the models getting cheap fast. For a product built on top of them, that is a cost going down.

What a company still owns once the model costs almost nothing:

  • Audience and network effects. The people already there, and the pull each new one adds.
  • Format. A recognisable way of doing the thing that others end up copying.
  • The relationship with the user. Trust and habit, built one conversation at a time.
  • Data from real conversations. What people actually say and ask, which no competitor can buy.
  • Mutual understanding. An application that learns what a person likes to communicate and how. That shared context deepens with every exchange, and it becomes a moat.

What we look for

"One gives birth to two. Two gives birth to three. Three gives birth to the ten thousand things." Tao Te Ching, chapter 42.

The spoken word was one. The written word was two. AI is the third, and the three are giving birth to the ten thousand things. So we do not keep a sector list. The space is young and any list would be wrong within a year. These are examples of what fits, not the boundaries of it:

  • Creation through conversation. Tools where talking with another person is the act of creating, and the production is handled for you. Ahmad is currently exploring a personal angel investment in Convo, which sits in this space. Memo to follow.
  • Knowledge you can question. Books, documents and expertise you can speak with, that answer back and draw out what you already know.
  • Entry without reading. Products that let people who cannot read or type take part in work, learning and commerce.
  • Wisdom carried by people. Ways to capture and pass on what elders, experts and communities know, in their own voices.
  • The consumer becomes the producer. Products where the line between audience and maker blurs, and the time a person spends taking part is itself an act of creation.

One question we ask of every product: does it make people more capable over time, or more dependent?

The founders we back

We back resilient founders. Two backgrounds stand out: people who have built (or failed) a startup before, and people with deep experience in the sector they are building for, and/or someone who just doesn't know how to quit.

The founding team must show real autonomy and resilience: they make decisions without waiting for permission, and they keep going when things break. We look for evidence of both before we invest.

The digital mirror

Thirty years of technology that only worked when you interrupted your life.

Every screen demands the same thing: stop, look down, read, type. The people around you wait. Your hands wait. The digital world has only ever worked on its terms, and the terms are your eyes and your attention.

The first always-on devices are already here. A pendant that listens all day. A note-taker that sits in every meeting. A pair of glasses that hears what you hear. Most of them are clumsy, and some of them are unwelcome. But the direction is clear. The digital world is learning to listen to the analog one, instead of asking the analog one to type.

Follow that to its end. A person who never reads a screen could still search, learn, buy, sell and remember, because the digital world keeps a mirror of their spoken life and answers when asked. You would live in the analog world, and visit the digital one by choice. The elder's village, with a very good memory.

That is a retrieval of the oldest way humans ever organised themselves, around speech and around each other. It will change habits, expectations and what we build our days around, the way print did and the way the screen did. The companies that make it work well are what this thesis is for.

One line we hold. A mirror that listens must belong to the person it listens to. Recording without consent is surveillance, and we do not back it.

Why Ahmad

For founders.

Ahmad has founded products and small startups himself: an online registration and event technology business in 2012, and e-commerce apps for small and medium businesses from 2017 to 2021.

At Canva, a company growing at speed, he was the first product manager on generative AI and launched Magic Write, an AI writing tool, before ChatGPT and the first LLM chat interfaces. He has built user experiences which generated tens of millions of monthl active users, alongside engineers, designers and founders. He later ran Canva's business across the Middle East, North Africa and Pakistan as general manager.

Since then he has spent his time on one question: what the return of the spoken word means for how people think, learn and build. His essays on McLuhan, Ong, the laws of media and the three ages of orality are on this site. Founders who read them will know how he thinks about product strategy in this space, and that thinking is what he brings to a cap table.

Deal flow follows from the thinking. The founders we want to work with already see the return of the spoken word. This site is written for them.

For investors.

A fund on this thesis would be based in Dubai, a tax-advantaged jurisdiction. It would be run by someone who has founded companies and shipped AI products, not only invested in them. One study of venture returns found that investors who had founded companies had success rates of nearly 30%, against just over 23% for professional investors (source (opens in a new tab)).

What we will not back

  • Anything that disenfranchises people, cuts them out, or makes them feel smaller.
  • Products too technical for ordinary people to understand.
  • Surveillance.
  • Military and defence applications.
  • Voice cloning or deepfakes of real people without their consent.
  • Products that replace human connection rather than extend it. A companion that only talks back from text is secondary orality with a voice on top. Hybrid orality keeps real people in the loop.
  • Products aimed at children without strong safeguards.
  • Companies whose purpose is not, in plain terms, to make the world safer and healthier.

How a fund on this thesis would work

A sketch. There is no fund, and nothing here is an offer.

  • Size: $25 million.
  • First cheques: $250,000 to $1 million, in 10 to 20 companies.
  • Reserves: half the fund held back to invest again in the companies that are working.
  • Role: a meaningful early backer, investing alongside a lead rather than setting terms.
  • Timing: a five-year investing period. The fund can run longer to hold its winners, with secondary sales as one route to exit.
  • Fees: 2% a year to run the fund, and 20% of profits once investors have their money back.

How we could be wrong

  • Reading is faster than listening. Speech may win for creating and asking, and still lose for taking information in (though we believe this is where the visual and dashboard applications will stay strong).
  • People may not want to talk to devices in public, at work or around family.
  • Speech recognition is weakest in the languages the access leg depends on.
  • The largest platforms may own the interface. If Apple, Google and OpenAI keep the conversation layer for themselves, there is less room for new companies.
  • Voice prices may stop falling. Several providers have held their prices flat for a year or more.

What would change our mind: total minutes of AI voice no longer growing while the price per minute falls.

The number we track: total minutes of AI voice, set against the price per minute, over a five-year outlook.