How Tenstorrent Plans to Make Money From Open-Source AI Chips

Here’s a puzzle. Tenstorrent gives away its AI software for free. Yet investors just valued the company at $6.2 billion. So how does a chip startup make money when it shares its code with the world?

The answer is smarter than it looks. In October 2026, the company closed a $1.5 billion Series E round. Trace Capital and Innovation X led it, while Fidelity, ARK Invest, and Morgan Stanley also joined.

In fact, that’s more than double the $2.6 billion mark from its December 2024 round. The CEO is chip legend Jim Keller, who worked on chips at AMD, Apple, and Tesla. His pitch is short: “We built one architecture and sell it at every size.”

In this listicle, you’ll learn the five ways Tenstorrent plans to turn open-source AI chips into real revenue. Plus, you’ll see the risks, so you can judge the plan for yourself.

Tenstorrent Uses Free Software as Its Sales Funnel

First, let’s start with the part that seems to lose money. Tenstorrent puts its whole software stack on GitHub. This includes the TT-Forge compiler, the TT-Metalium SDK, and the TT-LLK kernel library. Furthermore, the low-level code uses the MIT license, so anyone can read, change, or reuse it.

So why give it away? Software is the most significant barrier to entry for Nvidia. CUDA has had about 20 years to build tools and habits. No startup can win that race with a closed copy. That’s why the company flips the game. It lets developers see every layer, with no black boxes.

How the Funnel Works

Think of it as a free trial with no time limit.

• First, a developer ports a model from PyTorch, JAX, or ONNX with TT-Forge.
• Then they tune it with TT-Metalium. Once the code runs well, they need hardware to run it at scale.
• And that hardware comes from just one place.

Better still, the “try it” step keeps getting easier. Futurum reports that about 90% of tested Hugging Face models now run on the stack. So most teams can test their models in days, not months.

The Hidden Cost Saver

Here’s an angle most articles miss. Open source does more than attract users. It also cuts costs. For example, the company runs a bounty program and an active Discord. Outside engineers fix bugs and add model support as a result.

In effect, the community shares the work of building the open-source AI software stack. For a startup competing against a giant, this process presents a major obstacle.

Keller has called open source a one-way street. In other words, once you open the code, you can’t close it again. That’s a risk. But it also builds trust, because buyers know they won’t get locked in.

What You Can Learn

If you run a tech startup, ask yourself one question. Which part of your product is the “door,” and which part is the “room”? Give away the door. Then charge for the room. That’s the core lesson from Tenstorrent’s model.

Selling AI Hardware at Every Price Point

The second engine is simple. The company sells hardware. However, it offers a range of sizes, from a desk to a data center.

Here’s the lineup, based on public list prices:

• Blackhole P100A card: $999
• Blackhole P150A card: $1,399
• TT-QuietBox 2 workstation: from $9,999 (4 chips, liquid-cooled)
• Galaxy Blackhole server: from $110,000 (32 chips, 23 petaflops FP8)
• Four-Galaxy supercluster: from $440,000

Why the Ladder Matters

• Each rung leads to the next. For instance, a student buys a $999 card to learn.
• Next, a startup buys a workstation to test.
• Later, a company places an order for a full server cluster.

Since the same software runs on every rung, code from a cheap card can, in theory, scale to a supercluster. As a result, moving up feels less risky.

The Galaxy Blackhole server became widely available in April 2026. The company says it runs DeepSeek R1 at over 350 tokens per second for each user. It also claims a lower cost per token than Nvidia’s GB300.

However, no outside lab has checked those numbers yet. So treat them as claims until third-party tests arrive.

The Memory Trick Behind the Margins

Now for a less obvious point. The company skips pricey high-bandwidth memory (HBM). Instead, it uses cheaper GDDR6 plus lots of on-chip SRAM. Keller once said, “You can’t beat Nvidia if you use HBM.”

Because HBM supply is tight, this choice protects both price and supply. In short, this allows Tenstorrent to undercut rivals while still maintaining a margin.

Keller has also shared his profit view. He said that with a margin of about 10%, you struggle to make money. Yet above 50%, you invite rivals in. Therefore, you should anticipate that it will be priced in the middle range.

In other words, the company aims to position itself as an affordable AI accelerator rather than a luxury option.

A Tip for Buyers

If you’re an IT leader, start small. First, buy one card and test your top three models. Then compare the cost per token with your current GPU bill. Only after that should you think about a Galaxy order.

Licensing Tenstorrent’s RISC-V and AI Chip Designs

Here’s where the model gets clever. The company doesn’t only sell chips. It also sells the designs inside them. In this case, other firms pay to put Tenstorrent’s blueprints into their chips.

Two products lead this business. TT-Ascalon is a fast RISC-V CPU core. Meanwhile, Tensix is the company’s AI core. In June 2026, it also added TT-Ascalon S, a smaller version.

Named licensees include Samsung and LG. For example, LG has worked with the company on chips for smart TVs since 2023. Hyundai also takes part in its car chip plans.

Open Standard Is Not the Same as Open Source

Many readers get stuck here. If RISC-V is open, what is there to license? Actually, the answer is simple. RISC-V is an open standard, like a dictionary of words. Anyone can use it for free.

But the design that runs those words belongs to whoever builds it. The company spent years on its design. So that design is the product itself.

Why RISC-V IP Licensing Pays Well

Licensing is a high-margin business. After all, there’s no factory, no stock to store, and almost no cost per sale. Plus, it can bring royalties for years as customers ship chips. That’s why some analysts see it as the most valuable engine.

Tenstorrent also differs from Arm. Arm mostly sells closed “black box” cores. By contrast, the company offers source-code licensing. This means customers can see and change the design. For a company like Samsung, that means real ownership.

The Catch

Still, licensing is slow money. Counterpoint Research notes that this revenue depends on customers’ chips reaching production. And that can take years. Also, it faces strong rivals here, such as Arm and SiFive.

Our Take: Two Engines Beat One

Here’s the hidden strength. Licensing works as a hedge. If server sales grow slowly, licensing can still grow with every custom RISC-V chip. On the other hand, if licensing stalls, hardware still brings in cash. So two engines lower the odds that the whole business stalls at once.

Co-Designing Chiplets for Countries That Want Chip Independence

The fourth engine targets a new kind of buyer: governments. Today, many nations want AI chips they control. They prefer to diversify their suppliers instead of relying fully on one US supplier. Naturally, the company is eager to help.

The Japan Case Study

The best example is Japan. In February 2024, Japan’s Leading-Edge Semiconductor Technology Center (LSTC) picked Tenstorrent for a new 2 nm edge AI chip. The deal has several layers:

1. First, LSTC licenses Tenstorrent’s Ascalon RISC-V CPU tech.
2. Next, the company co-designs the chip with LSTC.
3. Then, it supplies RISC-V, LPDDR6, and I/O chiplets.
4. Finally, Japan’s Rapidus will make the chiplets on its 2 nm process.

So one deal brings in license fees, design fees, and chiplet work. That’s three revenue streams from one customer. On top of that, a Tenstorrent spokesperson said similar talks were under way in Europe and Korea.

Why Chiplets Change the Math

Chiplets are small chip blocks that snap together in one package. As a result, a customer can mix parts from different makers. For a new chip company, this opportunity is huge. Instead of building a full chip, a buyer can license ready-made blocks.

Counterpoint says the cost and risk of a first chip is the hardest part for chip startups. Tenstorrent’s chiplet licensing speaks right to that pain.

Selling Independence, Not Just Silicon

Here’s what many reports miss. In these deals, the company isn’t really selling chips. Instead, it’s selling sovereignty. A government pays for the right to own its silicon, tweak it, and build it at home.

Nvidia can’t easily make that promise, because its model depends on closed control. And as sovereign AI budgets grow, this sector could become a steady stream of large contracts.

What to Watch

However, these deals depend on partners. For example, Rapidus still needs to prove it can make 2nm chips in volume. If a factory slips, the whole project slips too.

Scaling Through Data Centers and Cloud Partners

The final engine is scaled. The company wants its Galaxy servers in data centers around the world. Yet it avoids one costly trap. It doesn’t try to build a giant cloud of its own. Instead, it lets partners do that.

Real Deployments So Far

The list keeps growing. For example, Equinix set up 16 Galaxy systems at its Ashburn hub. Furthermore, Cirrascale offers Blackhole cloud access, and Koyeb has offered Wormhole instances.

In India, Turiyam began a first phase of up to 32 galaxies. Meanwhile, Japan’s AI has the largest reported setup, with 120-plus systems. Keller has also mentioned a 96-Galaxy order headed outside the US.

How the Series E Fits In

The new $1.5 billion has three main uses.

• First, scale up Galaxy making and rollout.
• Second, keep serving open-weight language and video models.
• Third, build next-generation CPU and AI designs for licensing.

In short, the money feeds every engine at once.

Why Partners Beat a Private Cloud

Building a cloud costs a fortune. Worse, building a cloud turns you into a rival to your customers. By selling through partners, Tenstorrent keeps its balance sheet light.

That way, partners take the risk of filling the racks. Meanwhile, the company focuses on chips, software, and support.

The Risks Are Real

Still, stay calm and focused. The firm is private and doesn’t share revenue. In late 2024, it reported about $150 million in customer contracts. That’s tiny next to Nvidia’s data center sales.

Furthermore, a two-year chip cycle means one delay leaves a long gap. And the software gap with CUDA hasn’t fully closed.

Steps to Track the Tenstorrent Business Model

• First, track new IP license names after Samsung and LG.
• Next, watch for repeat Galaxy orders, not just first orders.
• Then, look for outside tests on cost per token.
• Finally, follow IPO news, since Keller has said the company is aiming for one.

Conclusion: Can Tenstorrent Turn Open Source Into Profit?

So, how does Tenstorrent plan to make money? It runs five linked engines.

• First, free open-source software pulls developers in.
• Next, hardware sales turn those users into buyers, from a $999 card to a $440,000 cluster.
• Then IP licensing lets firms like Samsung and LG build Tenstorrent designs into their own chips.
• On top of that, chiplet and co-design deals sell chip independence to nations like Japan.
• Finally, data center partners scale it all without heavy cloud costs.

The big idea is simple. Tenstorrent gives away the door and charges for the room. Open code builds trust and a community. Meanwhile, hardware and IP bring in the cash.

Of course, nothing is certain. Revenue is still private, benchmarks need outside proof, and Nvidia’s lead is huge. But with $1.5 billion in fresh funding and a $6.2 billion valuation, investors clearly believe the bet can pay off.

So here’s your next step. If you’re a developer, grab TT-Metalium from GitHub and try one model this week. If you’re a founder or investor, study how this model mixes free and paid layers. Then ask how you could use the same idea in your own business.

FAQs About How Tenstorrent Makes Money

1. How does Tenstorrent make money if its AI software is open source?

Tenstorrent gives its software away to win developers. Then it earns money by selling AI cards, workstations, and Galaxy servers. It also licenses its RISC-V CPU and AI chip designs to firms like Samsung and LG.

2. What AI chip IP does Tenstorrent license to Samsung, LG, and others?

Tenstorrent licenses its TT-Ascalon RISC-V CPU cores and its Tensix AI cores. Customers build these designs into their own chips. Unlike Arm’s closed cores, Tenstorrent offers source code licensing, so buyers can study and change the design.

3. How much does a Tenstorrent Galaxy Blackhole AI server cost today?

A single Galaxy Blackhole server starts at about $110,000 with 32 Blackhole chips. A base supercluster of four servers starts at $440,000. For smaller budgets, Blackhole PCIe cards start at $999 for developers.

4. Why does Tenstorrent use RISC-V and not Arm for its AI chips today?

RISC-V is a free, open standard, so Tenstorrent pays no Arm license fees. Because of that, it can build and license its CPU designs. It also gives customers more control and less vendor lock-in.

5. Is Tenstorrent a public company, and can investors buy its stock yet?

No, Tenstorrent is still private. It raised $1.5 billion at a $6.2 billion valuation in October 2026. Keller has said the company is aiming for an IPO, so public investors may get access later.