Avatar: The Way of Water has an extended underwater sequence that runs for roughly 30 minutes. It is one of the most technically complex sequences ever put on screen.

Here is how it was made.

The actors were real. Kate Winslet, Sigourney Weaver, and the rest of the cast spent months learning to free-dive so they could perform full scenes underwater without breathing equipment. Winslet held her breath for 7 minutes and 14 seconds during one take. The production team built a tank 120 feet long, 60 feet wide, and 30 feet deep, holding 250,000 gallons of water. The actors wore suits covered in tracking dots. Cameras captured every movement of their bodies and faces while they were submerged. That footage, the actual human performances, is the base layer.

To transform that footage of a group of actors in suits into Na'vi characters living in an underwater world, a second production takes place entirely in computers.

Animators use tracking data from dots on actors' suits to drive digital characters. Every movement Kate Winslet made underwater gets translated into the movement of her Na'vi character. Every facial expression gets mapped onto a digital face with different proportions, different skin, and different eyes. Human performance becomes the skeleton within a completely different body.

At the same time, a separate team of artists builds the world that those characters will inhabit. Every piece of coral. Every creature. Every shaft of light coming through the water. Each of these is a digital object with its own geometry, texture, and physical properties. The water itself has to behave like water: moving, refracting light, casting shadows, and interacting with the characters swimming through it.

When all of those elements exist, they have to be assembled into a single scene and calculated into a final image. That calculation is rendering.

The computer processes every pixel in the frame and determines what color it should be based on where each light source is, how that light travels through water, which surfaces it hits, and how it bounces. It does this for every pixel. In every frame.

A single complex frame in a sequence like this takes between 4 and 10 hours to render on one machine. Avatar's underwater sequences ran at 48 frames per second. Five minutes of that footage is 14,400 frames. On a single computer, rendering five minutes of the underwater sequence would take between 57,600 and 144,000 hours. Between 6 and 16 years of continuous computing.

That is the problem render farms solve.

A render farm is not a single powerful computer. It is hundreds or thousands of computers all working on the same job simultaneously. Instead of one machine working through frame 1, then frame 2, then frame 3 in sequence, a render farm assigns different frames to different machines at the same time. With 1,000 machines running in parallel, a job that would take 16 years on a single computer is completed in weeks.

Weta FX in New Zealand operates one of the largest render farms in the world. So does Industrial Light and Magic. Building and running that infrastructure costs hundreds of millions of dollars. Access to it determines who can make films that look like Avatar.

Avatar: The Way of Water spent $250 million on visual effects. Of that, roughly 60 to 70 percent went to the artists, animators, and supervisors doing the creative work. The remaining 30 to 40 percent, somewhere between $75 and $100 million, went to compute costs: the render farms, software licensing, and the infrastructure required to run them.

A single effects shot on a comparable film like Alice in Wonderland costs $46,000. Hollywood blockbusters typically spend between 20 and 40 percent of their total production budget on visual effects. For a $200 million Marvel film, that can mean $80 million or more.

Weta FX built and operates the infrastructure that rendered Avatar. Industrial Light and Magic, owned by Lucasfilm, runs the infrastructure for Star Wars and Indiana Jones. These facilities are private infrastructure owned by the companies that can afford to build them.

The same constraint exists at every level below Hollywood. An independent animator working on a short film needs the same type of processing power, just less of it. A game studio building a cinematic trailer, an architect visualizing a building, a product designer rendering a prototype — all of them need GPU compute they either cannot afford to own outright or do not need consistently enough to justify the expense.

The current options are to buy expensive hardware that sits idle most of the time, or rent from Amazon, Google, or Microsoft at prices and availability those companies control.

Think of it like accommodation before Airbnb. Hotels controlled all professional accommodation. If you needed a room, you paid hotel prices, on hotel availability, under hotel terms. Airbnb did not replace hotels. They still exist and still dominate. But Airbnb created a parallel marketplace where people with spare rooms could rent them to those who needed them, entirely outside the hotel system. GPU compute has no equivalent marketplace yet. Render is trying to build one.

The GPU compute problem did not start with AI. It started during the COVID pandemic, when supply chains broke down, and demand from gamers and crypto miners overwhelmed manufacturing capacity. By 2021 and 2022, graphics cards were selling at two to three times their retail price on secondary markets. The VFX industry was already competing for render capacity it could not reliably access.

Independent animation studios were building waitlists. Freelance artists priced jobs based on computing availability, not just creative time. Small studios were making production decisions based on what they could afford to render, not what they wanted to make.

That was the baseline before AI arrived.

Training an AI model requires the same type of processing power that renders a film, but at a scale that makes Hollywood look modest.

The largest AI companies run GPU clusters continuously, not for months during a production cycle, but permanently. Microsoft, Amazon, Google, Meta, and Oracle are collectively projected to spend over $700 billion on AI infrastructure in 2026 alone. That demand has collided with a manufacturing system that cannot keep up.

Lead times for Nvidia's most advanced chips are currently 36 to 52 weeks. SK Hynix, one of the primary memory chip manufacturers that AI processors require, has sold its entire 2026 output. Jensen Huang said in Nvidia's most recent quarterly results that cloud GPUs are sold out. Independent analysis suggests the supply gap will not fully close until 2028 or 2029 at current investment rates.

The organizations that locked in long-term cloud contracts in 2024 and 2025 have a meaningful advantage. Everyone else is competing for what remains, at prices set by the companies that own the infrastructure.

Jules Urbach founded OTOY in Los Angeles in 2008. The company he built produces OctaneRender, the rendering software used to produce visual effects in Westworld, Avatar, and multiple Marvel films. When Apple showcased the graphics capabilities of its M-series chips at its developer conference, it used OctaneRender as the demonstration software.

Working inside the professional rendering industry gave Urbach a direct view of the same problem from both sides.

On one side, artists and independent studios are priced out of serious rendering work. The computer they needed existed. They simply could not access it at a price that made sense. A freelance animator could not rent time at Weta FX. A small game studio could not afford to build its own render farm. The barrier was not creative ability. It was infrastructure access.

On the other side: enormous amounts of GPU compute sitting idle. Gaming computers running at partial capacity. Small studios with hardware that went unused overnight. Individuals with high-end graphics cards are doing nothing most of the day.

Urbach filed one of the first patents for decentralized GPU computing in 2009. He spent the next eight years building OTOY and OctaneRender, establishing the technical credibility and industry relationships that would make a marketplace viable. In 2017, he formally launched Render Network. The public mainnet went live in April 2020.

The idea was straightforward: connect idle GPUs with artists who need them, handle payments automatically, and take a fee on each transaction.

Render is a marketplace. GPU owners register their hardware with the network and make it available to rent when they are not using it. Artists and studios submit rendering jobs through the network, pay for the compute they use, and receive the finished output.

The payment mechanism is designed to solve a specific problem. Creators pay in dollars, not crypto. The network converts that payment into RENDER tokens and burns them, removing them from circulation. It then issues new RENDER tokens to the GPU owners who completed the work. Creators do not need to hold or manage crypto to use the service. And when more jobs run, more tokens are burned, creating a direct relationship between actual network usage and token supply.

Before a node operator gets paid, the work is verified. Every node is benchmarked upon joining, assigned a performance score based on its hardware, and ranked accordingly. Higher-scored nodes get priority on jobs. Creators review completed renders within 24 to 48 hours and approve them before payment is released. Nodes with stronger completion histories build reputation scores that affect how much work they receive going forward.

As of 2025, the network has processed over 63 million rendered frames, 22 million of them in 2025 alone. There are 15,000 active nodes. A digital artist named Brilly used Render for a Coca-Cola campaign displayed on the Las Vegas Sphere. Andy Torres used it for the Super Bowl LIX countdown trailer. Animation studio V! Studios rendered over 100 jobs on the network for NASA's Benefits for Humanity program, producing 4K videos about the International Space Station. These are production jobs for real clients with real deadlines.

In 2025, Render launched a platform called Dispersed specifically for AI compute, aggregating its distributed GPU network for AI inference and training workloads. The network has partnerships with Stability AI and Luma Labs, both of which use GPU compute for AI-generated imagery.

Render passed my Gate and Key evaluation.

The use case is real, and the problem it solves existed before AI made it urgent. The founder has a 15-year operating track record in the exact industry the network serves. The infrastructure is functional and verified, and it has processed tens of millions of jobs. The creative community adoption is genuine and sticky. The expansion into AI compute is the right strategic direction given where GPU demand is going.

The primary risk is the gap between current revenue and current valuation. The structure is sound. Whether the market captures what the price is pricing in is an open question, which the math section below addresses.

If you own a qualifying Nvidia GPU, you can apply to become a node operator.

The minimum hardware requirements are an Nvidia GPU from the RTX 3050 series or above, at least 6GB of video memory (8GB or more recommended), 32GB of system RAM, 100GB of free storage, and a reliable internet connection. Render's own documentation is clear on one point: do not buy hardware specifically for this. The network is designed for people who already own qualifying hardware and want to put idle capacity to work.

What you would earn depends on your GPU's performance score, how much time it is available, electricity costs in your location, and the volume of jobs the network is processing. In the first half of 2025, the network distributed 462,115 RENDER tokens to node operators across all active nodes. Income is variable, paid in RENDER tokens, and moves with both network demand and token price.

There are no documented individual case studies of specific people earning a full living from Render node operation in the public record. What exists: 15,000 nodes are active on the network, which means real people have found the economics worth their time. Render's Discord community and rendernetwork.com are the most reliable starting points to understand what participation looks like for your specific hardware.

Render currently earns $4.8 million in annual revenue. The GPU-as-a-service market is worth $8.21 billion today and is projected to reach $26.62 billion by 2030, growing at roughly 27 to 35 percent annually. That puts Render's current market share at approximately 0.06 percent.

Here are the two honest scenarios.

If Render grows with the market but gains no additional share:

At 30 percent annual growth, revenue reaches roughly $18 million by 2030. Against a current valuation of $868 million, that requires a 48x revenue multiple just to stay flat. In this scenario, the valuation compresses significantly.

If Render captures 0.5 percent of the 2030 market:

0.5 percent of $26.62 billion is $133 million in revenue. That is 28 times current revenue. At a 20x multiple, that gives a 3x return from today. At 30x, a 4.6x return.

For a 5x return: revenue needs to reach $217 million, which equals 0.8 percent of the projected 2030 market. That is a 13x increase in market share from today.

For a 10x return: revenue needs to reach $434 million, which equals 1.6 percent of the 2030 market. A 27x increase in market share.

The hotel market Airbnb entered grows at 6 to 9 percent annually. The ride-sharing market that Uber entered is growing at 13 to 18 percent. GPU compute-as-a-service growth is 27 to 35 percent. Airbnb and Uber competed for a relatively stable pool of buyers. GPU compute demand compounds. Every new AI model requires more of it. Every application built on those models requires more to run.

When Airbnb had roughly $5 million in revenue, investors could only access it through a private venture round at a $70 million valuation. Uber was the same. Both were inaccessible at that stage. Render is at $4.8 million in revenue today, and anyone can buy in. The valuation is $868 million, significantly higher than either at the same revenue stage. That premium is the price of accessibility in a market that those two companies never had to compete in.

TLDR

Render holds 0.06 percent of an $8.21 billion market, growing at 30 percent annually.

Growing with the market will get revenue to $18 million by 2030. The current valuation is very hard to justify at that level.

Capturing 0.5 percent of the 2030 market gets revenue to $133 million. At a 20x multiple, that is a 3x return. At 30x, a 4.6x return.

A 5x return requires 0.8 percent market share. A 10x return requires 1.6 percent.

The $4.8 million is the starting line. The bet is on how much of a tripling market this network captures over the next five years, in a category where demand has no ceiling in sight.

Revenue relative to valuation. At $4.8 million in annual revenue against an $868 million valuation, the network needs significant usage growth to justify its current price. Growing with the market is not enough. Real market share capture is required.

Competition. Akash Network offers general-purpose compute on similar principles. io.net focuses specifically on GPU clustering for AI workloads. Amazon, Google, and Microsoft have more reliability and deeper enterprise relationships. Render's advantages are its established creative community, its OTOY software foundation, and 15,000 active nodes that represent real supply-side infrastructure.

Node operator economics are tied to the token price. Earnings for GPU owners fluctuate with both network demand and the RENDER token's dollar value. In late 2025, RENDER's price dropped 76 percent from its peak while network usage hit record levels. Node operators were processing more jobs for less dollar-denominated income. That tension is worth understanding before participating on the supply side.

The revenue figure has a structural limitation. Part of it is calculated by converting RENDER tokens earned by contributors at the current token price. When the token price falls, reported revenue falls even if network activity is unchanged. Targon's $10.5 million Series A, raised in conventional dollars from outside investors, is a more reliable signal of real economic value being created.

Three things.

Size the position for total loss. This is a speculative bet on a network capturing market share in a new category. No prudent investor puts money here they cannot afford to lose entirely.

Build the position over time. If the thesis is correct, buying in stages lets you test the evidence before committing fully.

Define the thesis with specific measurable milestones before you buy, and commit to exiting if those milestones are not met.

What to watch over the next 12 to 18 months:

Monthly RENDER token burns should be increasing. Burns are a direct proxy for network activity. Flat or declining burns in a growing AI compute market mean the network is losing ground.

Active node count should be growing meaningfully above 15,000.

Dispersed revenue should be emerging. The AI compute expansion launched in late 2025. By mid-2026, there should be early data on whether AI workloads are actually flowing through the network.

The GPU shortage should persist. Current projections indicate it will continue through 2027-2028. If a new chip supply resolves it faster, the urgency for an alternative marketplace weakens.

Competitors Akash and io.net should not be visibly pulling ahead on node count, job volume, and enterprise adoption.

The honest verdict:

At the current price, this is a speculative bet that a network holding 0.06 percent of its addressable market will capture 0.5 to 1.6 percent of a fast-growing category within five years. The structural conditions for that to happen exist. The infrastructure is real. The founder has 15 years of industry credibility. The market is growing faster than almost anything else in technology.

But the price already reflects optimism, not current reality. A prudent investor treats this as a small, monitored position with a clear thesis and a clear exit trigger. Not a conviction bet.

The next 18 months will start to provide an answer.

This is a structural analysis. Nothing here is financial advice.

Next up: Arweave.