When we think about the application layer of LLMs, it is easy to think of practical vertical use cases such as software development, legal, or finance, but the top use case of Generative AI is actually in the realm of Therapy and Companionship. The demand for companionship is not new, but the technology behind it has been moving quickly in the last few years. A clear line runs from Twitch to VTubers to AI companions. Each step captured demand the prior format couldn’t reach. Shizuku AI, founded by Akio Kodaira in August 2025 and backed by a16z at a $15M seed at $75M valuation, is aiming to build the next generation of technology through the form of AI VTubers.

Founding Story

Shizuku AI the company was founded by Akio Kodaira (CEO) in August 2025. Kodaira launched an AI VTuber named Shizuku on YouTube while still completing his Ph.D. at UC Berkeley in 2023. With Shizuku he ran dozens of livestreams and built a community of thousands of followers. On the research side, Kodaira was a lead author on breakthrough papers such as StreamDiffusion showing real-time image generation fast enough for live video applications. Kodaira also worked on real-time video generation at Meta and Luma AI, giving him the technical edge to compete at the frontiers.

Evolution Of Technology

Although we could consider the field Shizuku AI is playing broadly as Entertainment, I'd like to narrow down on this idea of companionship and explore the historic evolution of companionship and the technologies surrounding it to the present day. Thinking from first principles, companionship as a service is the experience of focused, reciprocal attention from another mind delivered through a medium in exchange for value. Back in 1956, psychologists Donald Horton and Richard Wohl coined the term parasocial relationship to describe the illusion of intimacy TV viewers developed with on-screen personas. We see this idea persist as a 2017 Google study found that 40% of millennial Youtube subscribers said their favorite creators understood them better than their friends. There seems to be an underlying model where emotional investment in a media figure has turned into an economic engine that powers a multi-billion dollar human livestreaming industry and a rapidly growing virtual character industry. We can observe the history of making focused attention cheaper, more scalable, and more accessible to make a reasonable prediction of how companionship technology will evolve. Each wave of technology can somewhat say they replaced the real thing with some type of artificial intimacy, yet each wave is adopted much faster than the incumbents predict as it serves latent demand that could not be reached previously.

Twitch

Twitch launched in June 2011 as a gaming live-stream platform and quickly became one of the largest companionship platforms in the world. Twitch had the parasocial dynamics of broadcast media, but also had a direct payment layer through subscriptions, donations, and bits, allowing viewers to purchase these micro-moments of attention from a streamer in real time. When we look at some of the numbers from Twitch it is quite significant.

• 240 million monthly active viewers as of early 2025, up from 55 million in 2015.

• 20.8 billion hours of content consumed in 2024, with a peak of 24.3 billion hours during the pandemic era.

• $1.8 billion in estimated revenue in 2024, generated through subscriptions (58%), advertising (33%), and Bits/virtual goods (9%).

• Just Chatting became the #1 most-watched category on Twitch in 2023, accumulating 2.86 billion hours of global watch time.

While the MAUs and revenue figures are impressive, Just Chatting being the #1 most-watched category didn’t make sense intuitively, but it accumulated more watch hours than any individual game, including League of Legends (1.23 billion hours). It is important to note that Twitch tracks each game as its own separate category, while Just Chatting is one giant umbrella. Even with this in mind, the demand for unstructured content that fits this category of companionship is notable. Researchers described the Twitch parasocial relationships as one-and-a-half-sided because Twitch offers intermittent moments of genuine reciprocity when a streamer reads a donation or responds to a chat message. The possibility, but not certainty, of being noticed creates what behavioral psychologists recognize as the wedge to driving sustained engagement.

There are some limitations in the Twitch companionship model as it is one-to-many(the streamer cannot give sustained individual attention), time-bound (streamers sleep, burn out, and take breaks), and expensive to produce (top streamers are scarce human talent with leverage).

VTubers

In late 2016, a character named Kizuna AI debuted on YouTube. She was a 3D animated anime character and coined the term “Virtual YouTuber”. Within ten months, she had over 2 million subscribers. Her success triggered an industry: by January 2020, there were over 10,000 active VTubers. YouTube’s 2020 Culture and Trends report highlighted VTubers as a notable trend, with 1.5 billion views per month by October 2020. The VTuber model decoupled the companion from the physical person. The real human still performs behind the avatar, but the audience bonds with the character, not the individual. This had a few significant implications as characters became ownable IP. The VTuber character belongs to the agency. Cover Corp (Hololive) and AnyColor (Nijisanji) own the characters their talent performs. If a performer leaves, the character can, in theory, continue with a new voice. These characteristics of this new domain have made the VTuber market quite lucrative.

• $3+ billion global VTuber economy in 2025.

• Cover Corp (Hololive): ¥13 billion (~$80M USD) revenue in FY2026/Q3, 88 managed VTubers, 80M+ combined subscribers.

• AnyColor (Nijisanji): ¥42.9 billion (~$286M USD) revenue in FY2025, 34% YoY growth, 170 managed VTubers.

VTubers proved that audiences will form deep emotional bonds with an artificial persona. The human behind the avatar is anonymous and the audience attaches to the character. But VTubers still require a human performer for voice, improvisation, and emotional responsiveness. The performer remains the bottleneck.

AI Companions

Running parallel to the VTuber explosion, a separate category emerged: AI companion apps that replace the human entirely and restore the one-to-one intimacy that live streaming sacrificed for scale.

• 337 active, revenue-generating AI companion apps worldwide, with 128 released in 2025 alone.

• 82 million generated in H1 2025, on track for $120M+ for the full year, representing 64% year-over-year revenue growth.

• Character.AI: 20M MAU, 180M+ monthly website visits, 10 billion messages/month, $32.2M revenue in 2024.

• Replika: 30M+ registered users, 85% report emotional connections, $24M revenue in 2024.

Despite strong growth, current AI companions have a fundamental gap because they are text-based, reactive, and lack the performative, visual, communal dimension that makes VTubers and streamers engaging. There is no character to see, no live stream to gather around, no community to belong to. A16z, in their investment thesis for Shizuku AI, identified this as “the fundamental challenge facing AI companions today”, where monotonous, reactive conversations fail to sustain long-term engagement. Users eventually churn because the AI has no presence.

AI VTubers

The AI VTuber is one plausible evolution of the technology we have evaluated so far. A format that combines the visual character identity and communal experience of a VTuber with the always-on availability and one-to-one capability of an AI companion. It streams live, responds to viewers in real time, has a face and a voice to bond with, and never sleeps.

One example of an AI VTuber is Neuro-sama, created by the pseudonymous developer Vedal. Neuro-sama streams autonomously on Twitch, where it is driven entirely by AI, with no human performer, and has achieved extraordinary traction.

• 343,000+ peak Twitch subscribers in January 2026, making the channel the third most-subscribed in Twitch history behind only names like KaiCenat.

• Estimated seven-figure annual revenue from subscriptions alone, with peak monthly subscription revenue of approximately $400,000, before accounting for donations, Bits, sponsorships, and ad revenue.

Neuro-sama’s fans are said to engage in real-time co-performance where they actively shape the AI’s behavior through chat in the moment, rather than passively consuming content or remixing it after the fact. This creates a qualitatively different relationship than either traditional streaming or text-based AI companions.

Shizuku AI is building at this intersection. The character speaks Japanese and English, sings, and interacts with live-stream viewers in real time using LLM-driven conversation and text-to-speech synthesis.

Shizuku’s success hinges on solving the key failure mode of current AI companions. They are reactive, monotonous conversations that cannot sustain engagement. Shizuku AI’s approach is to deploy the character across YouTube, Discord, and X simultaneously, building a community and a data flywheel where real interaction data trains increasingly proactive, engaging conversational models. Shizuku is also a16z’s first Japanese investment, and this choice seems very intentional. Japan has the world’s largest VTuber industry (home to Cover Corp and AnyColor), world-class AI and animation engineering talent at a fraction of Silicon Valley costs, and an aging, increasingly isolated population that creates structural demand for exactly what Shizuku offers.

Kodaira himself is the rare founder who has operated on every layer of this stack. He built a live AI VTuber with a real community, co-authored StreamDiffusion to push real-time generation to the speed live interaction demands, and shipped AI systems at Meta and Luma AI. He has tons of experience building the infrastructure other people are now trying to use.

Every wave we analyzed from Twitch →VTubers → AI Companions, the incumbents called it a poor substitute for the real thing, and it still found a way to serve demand the previous format couldn’t reach. Is the next wave here? If you’re building in this space, I want to hear from you. If you’re interested in joining Shizuku, they are also hiring.

Munger argues that people fail to explain Coke’s success because they ignore the fundamentals. He boils it down to creating and maintaining conditioned reflexes. For Coke to dominate, it simply needed to maximize beverage ingestion and minimize the chance that those reflexes would be extinguished by competitors.

That chapter inspired me to look at the current star of the AI world, Cursor, and try to find their fundamental driver of success.

I combined Munger’s thinking with a framework on legibility to capital from Will Manidis (Founder of ScienceIO) and the recent Inside Cursor dispatch from Colossus. When I merged these three inputs, it reminded me of the current NBA favorite, the Oklahoma City Thunder.

General Manager Sam Presti has built the Thunder into a contender by sticking to a very specific style of asset management and culture. When I read the details of how Cursor operates, I realized they are effectively running the Presti playbook.

The Sam Presti Mental Model

In hindsight, we can all say the OKC Thunder made all the logical decisions in their road to building a dynasty. However, to understand and articulate exactly how they did it, we must examine the three pillars of the Presti Model (my best Charlie Munger impression).

Asset Optionality → They continuously flipped current positive assets for a larger sum of future positive assets.

Positionless Systems → Instead of rigid roles, every player on the roster can pass, handle the ball, and shoot.

Legibility to Capital → The entire organization understands the goal, and is perfectly incentivized to move towards that goal.

That third point is the most important because it captures what Will Manidis recently wrote about companies becoming legible to capital.

As Manidis says, a firm is legible to capital when “by dint of their existence capital forms behind them in excess.” But this only happens when the trade is clear. He uses the chocolate cake problem to explain this alignment. Many firms have great inputs like eggs and chocolate, but management thinks they are cooking a soufflé, investors think they are getting cupcakes, and talent thinks they are getting a pound cake.

The companies that are most legible to capital are obsessive about chasing the tolerances between competing ideas of the firm to zero.

This is the secret of the OKC Thunder. Presti, the coach/owners, and the players were all baking the exact same cake. They were willing to lose for years and accumulate assets in order to build a dynasty because there was zero tolerance for competing ideas.

Finding the OKC Thunder of AI Startups

So if we take this blueprint of hoarding talent assets, positionless systems, and extreme cultural alignment, and look for the OKC Thunder of the AI world, we first have to look past the teams buying championships.

We have to ignore the Lakers or Warriors of AI like OpenAI or Anthropic. They are massive and expensive and filled with veterans or researchers paid like NBA superstars. OpenAI now has over 3,500 employees and wins through brute force and luxury tax spending on compute.

We are looking for a team that wins with greater efficiency. When you look at the market through that lens, one company stands out, and it is Cursor.

They recently became the fastest software company in history to hit $1B in ARR (they got to $100M with zero marketing spend). While the establishment relies on a massive payroll, Cursor still operates with a tight rotation of roughly 250 people.

Here is why Cursor is the Thunder.

1. The Draft Strategy ←→ Hoarding Atomic Units

The Thunder didn’t build their core by signing free agents to fill specific roster holes. They didn’t say “we need a center.” They drafted unicorns which are long and high-IQ players who could do everything and figured the fit out later. They played the numbers game.

Cursor does the exact same thing. Most startups still hire certain roles. They need a senior react engineer or a product marketer. Based on inside notes from the Cursor team, they operate differently. They treat the atomic unit of hiring as the person rather than the role. They have a slack channel where they just drop names of people who are outliers.

It doesn’t matter if there is a role open. If the person is talented they swarm them. They chased one engineer, Lukas Möller, for a year after he rejected their first offer. They hired another, Ian Huang, just because he was coding into the wee hours of the night, and they saw it in the customer telemetry.

This mirrors the Presti approach of asset optionality. They aren’t trying to fill a hole in the org chart. They are hoarding talent density. This results in a team where 50 employees are former founders. That feels like the NBA equivalent of a roster full of lottery picks.

2. The Positionless System ←→ The House of ICs

Having talented players is one thing, but the system in which they play is another. The Thunder play a 5-out offense that requires constant motion and decision-making from everyone. It creates a pace that slower and more structured teams can’t match.

In software the slower structured team relies on the assembly line where product managers write specs, designers make mocks, and engineers simply execute. It is super rigid and slow.

Cursor operates as a House of ICs (Individual Contributors). They have effectively removed the layer of middle management, so there is no hand-off. The roadmap is bottoms-up and driven by the people actually writing the code.

Inside Cursor described a recent weekend where a group of four locked themselves in a room to ship a major browser feature. No permission, no spec document, just pure execution. They really have that 'you can just do things’ mentality.

This is the operational version of positionless basketball.” By removing the friction of management, they increase the pace of play. They acquire more possessions and iterations than their competitors.

3. Alignment: Raising the Ceiling

This brings us back to legibility to capital. What is the cake Cursor is baking?

Most AI startups are trying to lower the floor. They want to democratize coding so anyone can do it. That is a messy and crowded market.

Cursor made the trade clear because they are raising the ceiling. They are building a tool to enable the top 1% of developers to become 10 times faster. They are betting on high-agency users. Because of this clarity, there is no air between the founders, the team, and the investors. The trade is immensely clear, which is why capital ($3B of it) formed behind them instantly.

Closing Thoughts

The reason the OKC Thunder are terrifying to the rest of the NBA isn’t just that they are good now. They are set up to be good for the next decade. They have one of the youngest rosters and the most draft picks.

Cursor feels the same. They aren’t winning because of a single feature that can be copied. They are winning because they built a machine that is better at building the product than anyone else. They have assembled the best talent, eliminated management friction, and aligned everyone around a common goal.

That is a dynasty in the making.

This past summer, I interned at KUSABI, an early-stage venture fund in Japan with a unique approach to investing. Unlike many funds in Japan that focus on "base hits" (safe, moderate returns), KUSABI's partners were committed to "home-run" investments, seeking out startups with the potential for huge impact. KUSABI was also structured in a rare way: each of its three GPs was individually responsible for a third of the fund's capital, with no investment committee to approve deals. This structure emphasized an actual "skin in the game" mentality, pushing each GP to take full ownership of their bets, a practice not common in Japan. These attributes about KUSABI attracted me when I considered which VC to work for over the summer.

Subscribe now

At KUSABI, Yoshida Junya, The GP I worked under, did not give me specific instructions on what to work on or how to contribute to the fund. All he told me at the start was: "Be different, do something big, be more than an intern." The first thing I worked towards was building my credibility and becoming someone founders could trust and seek advice. I needed to get my name out there in a relationship-driven industry like venture capital. After telling my plan to Yoshida, we agreed on a metric of gaining 1000 followers in one month, with his reasoning being that I am not resourceful enough if I cannot achieve this number.

In building a persona, I aimed to be "the guy who knows a lot about American startups," building a presence online and in person. I spent significant time experimenting with engaging potential founders and partners on social media, adjusting everything from post timing to hashtag choices, and trying to understand what drove more engagement.

I also took every opportunity to attend events, even volunteering as a waiter or receptionist, to meet new founders. As time passed, people began to see me as "the Y-Combinator guy" or "the orange guy," and it was easier to skip introductions and dive into conversations about their startups. This approach helped me successfully grow my following to over 1000 people in a month, and I saw a 20% increase in applications to the fund's accelerator program. This tweet below shows the formatting I settled on to portray information to my audience.

One of my hit products that increased my credibility was this YC startup database that I made.

Although Yoshida did not tell me much directly, being around him taught me many lessons. I will highlight the two most important lessons I learned.

The first lesson was to think about technology as a continuous timeline, where you should think about a specific product/industry's history and make various hypotheses on what the future of that product/industry will look like. Then, make a reasoning process explaining why one specific hypothesis is the best, and find a founder working on this problem. The other direction was to think about consumer behavior as a continuous timeline. It would apply the same thought process but from a consumer behavior/product adoption perspective.

To practice this way of thinking, Yoshida suggested building "mini-theses" through value-chain analysis. This involved tracing the history of a specific technology or market trend, mapping out where new value was being created, and identifying which parts of the value chain were most vulnerable or promising for innovation.

This kind of analysis forced me to segment and categorize each market, leading to insights into competition, potential value-adds, and market direction. As a simple example, Google's value chain changes drastically when considering the 35% fee it must pay Apple for its search product. This practice helped me form hypotheses about future market shifts, a mindset I think is vital to becoming a VC with a long-term view. The document below shows one example of this value-chain analysis.

The second lesson was that assessing founders was much more than their startup's pitch. One of the important questions he taught me to ask myself was, "Would I want to work for this person long-term, regardless of whether their startup succeeded?" If the answer was no, that was a signal about the relationship I'd be building with them. Venture capital, especially with a long-term outlook, is a long journey. Investing in someone you respect and trust isn't just a good idea but essential.

On a side note, Yoshida often preached to me during my time at KUSABI that the venture process hasn't fundamentally changed since Sequoia's early days in the 1970s: deal sourcing, due diligence, growth support, and exit planning still define the work. Yoshida pushed me to think about where there might be room to "hack" this process and develop a unique edge. I focused on deal sourcing and personal networking, building relationships that could continue to benefit the fund long after my internship. However, I must continue asking myself what hacking the other stages would look like because there is no reason to believe these processes shouldn't evolve in the future.

Overall, working at KUSABI shaped my view on venture capital and founders. It taught me to think long-term, to make thoughtful assessments, and to always ask how I could do more than expected.

Anyways, the CEO seemed to have taken a serious liking to me and said, let’s meet up to talk more this weekend. One meeting turned into another, and he taught me all sorts of things, from the best way to grow my network in Korea to how to craft an Informational Memorandum (IM) for one of his deals.

Making an IM was super challenging, as I didn’t previously understand the amount of due diligence required, spending hours researching the industry, the competitors, and the potential investors. As someone with very limited Korean, my first step actually started with using AI tools to translate all information from Korean to English and then to ensure the CEO that the translation didn’t miss any connotations; this was a very time-consuming process, but it definitely taught me the importance of patience.

After gathering the data and formulating reasoning, the most crucial part was the communication design. How can the document deliver all the critical messages without boring/distracting the reader? How can we ensure that the reader of this document understands and sees the same thing we envision? These are elementary things to do, yet many of us still fail. I am occasionally guilty of this, too, but building iterations of the document and using communication design principles for a meaningful real-world project was great.

As I reflect on my time in Seoul from September - December 2023, it is weird to think that I could form such a strong relationship from a random connection I made. However, I am super grateful for the experience and learnings that followed.

So, how did I get myself into this situation? I was on a sailing voyage full of High School Seniors who had previous sailing experience or were National Boy Scouts, and they, for some reason, thought that voting for me to be Captain so I was responsible for their lives was a good idea. In the previous 8 days, they must have seen something in the only foreign boy with no sailing experience that they would trust me.

Anyway, this story isn’t about how I got the opportunity but what I learned from this rare opportunity.

The TLDR to my learnings, if you don’t want to read the rest of this experience, was:

a. There is a particular way of leading a group of strong personalities with high expertise.
b. When emergencies arise, you just have to make decisions quickly and deal with them.


There is a particular way of leading a group of strong personalities with high expertise.

I believe effective leadership is very situational, and no particular style is correct. Charisma isn’t something you can practice; not everyone can be a strong-willed commander. I prefer to think of leadership as a puzzle. It is putting pieces together so that the stakeholders’ goals involved match, people are doing what they enjoy, and your people have significant trust in you to make a calm, logical decision. In my situation, my first Job as captain was to align everyone’s goals. I preached why we were all here; we gave up our New Year gatherings to experience something fulfilling. We were here to do challenges that would make us better people. Aligning our vision was crucial on our ship because the arrogant and experienced students suddenly became passionate and helpful managers. We then proceeded to do something the supervisor captain had never seen before in his 10 years running the program. The SILENT JIBE.

It requires everyone to coordinate to maneuver the masts simultaneously in order to catch the wind. I assigned 4 sub-leaders and sent out the decided hand signals as the wind was ready, BAM BAM BAM, just like that, we completed the first SILENT JIBE in Spirit of New Zealand history. This JIBE was the crucial turning point in our journey, as it instilled massive confidence in all of our members, breaking down barriers between skill levels and personalities. People thought I was ridiculous for suggesting this idea, but no one made fun of me after we succeeded. I guess when you have a group of EGOs, you have to set the goal even higher to keep these people challenged while taking the group's overall performance to the next level.

When emergencies arise, you just have to make decisions quickly and deal with them.

As it happens with everything else in life, things just have to turn south as soon as we start out good. The waves started getting super rocky, and the crew became seasick. People were dropping like flies, and after a few minutes, half of the crew were at the front of the ship trying to recover. Then suddenly, one of our lifeboats fly into the ocean. I took 2 seconds to process what happened……

I scream at the available crew to control the front two masts so we can make a hard right U-turn. We spend about 15 minutes in the rocky waters trying to hook the boat to pull it back up to our ship. This was definitely not on the menu when I initially planned the day. But I just reacted to the moment, and things turned out to be just fine. As the day passed, more members gradually recovered, and we reached our final destination. We ran into another situation near the end where we had to tightly navigate between ships trying to get into the harbor. Still, at this point of the day, I had seen so much that I could coordinate and build a quick strategy without any struggle.

My greatest achievement was not winning a math competition, hackathon, or pitch contest, but it was this very experience because I learned how to navigate challenging situations on the fly and bring people together.

For reference, here is a note from the captain of Spirit of New Zealand: