Intel in talks to acquire AI chip startup SambaNova, Bloomberg News reports

[Daniel Nenni]

U.S. chip-maker Intel Corporation is reportedly in early-stage talks to acquire AI-chip startup SambaNova Systems, according to sources cited by Bloomberg News. The potential acquisition would mark a strategic push for Intel into the growing artificial-intelligence hardware space, after previous efforts to roll out its own AI-GPU products lagged behind competitors.

SambaNova, which designs custom AI processors and systems, has been working with bankers to gauge interest from potential buyers. While the company achieved a valuation of around USD 5 billion in a 2021 funding round, any deal today would likely value it below that level. The discussions remain preliminary and there is no guarantee they will lead to a transaction.

Intel has longstanding ties to SambaNova: Intel’s venture arm, Intel Capital, is an investor in the startup, and former Intel executive and investor Lip‑Bu Tan serves as SambaNova’s executive chairman. Industry watchers view such a move as a way for Intel to accelerate its AI hardware roadmap and better compete in the data-centre and cloud-inference markets. The companies declined to comment on the report.

https://www.yahoo.com/finance/news/intel-talks-acquire-ai-chip-202208497.html

 

[KevinK]

Does the new LBT Intel really need and see a path with yet another also-ran AI architecture ? SambaNova did some smart things (rack optimized architecture, system software), but are carrying around the burden of 3 year-old chips when the industry is on a yearly cadence. The whole 16 bit “accuracy” is actually a liability for most LLMs today, though maybe an asset for HPC.

“SambaNova’s RDU excels at running very large, specialized models with high efficiency, in part due to the three-tier memory architecture mentioned previously. Independent benchmarks from Artificial Analysis measured SambaNova’s performance on the DeepSeek-R1 671B model at 198-255 output TPS. Note that this is 16 RDU chips, while Nvidia generates 250 tokens per second with just eight GPUs. albeit using FP4. SambaNova emphasizes its ability to run the full, unquantized (16-bit) version of these massive models on a single system, a feat it claims competitors achieve only by using smaller, less accurate model versions or significantly more hardware.”

Cerebras, Groq And SambaNova Line Up To Compete With Nvidia

These three companies are all delivering rack-scale infrastructure to compete with Nvidia's NVL72 class system for inference. How do they stack up?

www.forbes.com

 

[blueone]

I'm only guessing, but I suspect LBT may be buying a team more than a design.

 

[swka]

I'm only guessing, but I suspect LBT may be buying a team more than a design.

given the stalling market traction, it cannot be the product they are buying. what's the track record of the design team? the company have been a round for quite a few years, and there must be several generations of the design.

 

[siliconbruh999]

The track record of Internal M&A has not been great ...

 

[Daniel Nenni]

I'm only guessing, but I suspect LBT may be buying a team more than a design.

That is my understanding. As an early investor LBT knows this company intimately thus he knows the team, product, and market. Plus, there is a lot of back scratching in this business:

SambaNova Investors:

• SoftBank Vision Fund 2
• Temasek
• GIC (Singapore sovereign wealth fund)
• BlackRock
• Intel Capital
• GV (formerly Google Ventures)
• Walden International
• Redline Capital
• Atlantic Bridge Ventures
• WRVI Capital
• Celesta Capital
• 
  
• Funding History:

• Series A: ~$56 million, led by Walden International and GV; participation from Redline Capital and Atlantic Bridge Ventures.
• Series B: ~$150 million, led by Intel Capital; existing investors GV, Walden International, Atlantic Bridge Ventures, Redline Capital also participated.
• Series C: ~$250 million, led by BlackRock; participation from GV, Intel Capital, Walden International, WRVI Capital, Redline Capital.
• Series D: ~$676 million, led by SoftBank Vision Fund 2; new investors Temasek and GIC joined; existing investors BlackRock, Intel Capital, GV, Walden International, WRVI Capital continued. This round brought total funding to over $1 billion and valuation above $5 billion.

 

[osnium]

Does the fact that Tan is also the chairman pose a conflict of interest?

 

[blueone]

given the stalling market traction, it cannot be the product they are buying. what's the track record of the design team? the company have been a round for quite a few years, and there must be several generations of the design.

I don't know enough to say.

 

[blueone]

Does the fact that Tan is also the chairman pose a conflict of interest?

Was chairman, but the answer is yes, though it's not a showstopper. The classic example: Steve Jobs, Apple, and NeXT.

 

[ChrisGar]

Was chairman, but the answer is yes, though it's not a showstopper. The classic example: Steve Jobs, Apple, and NeXT.

According to Sambanova website, Tan is still the Chairman.

 

[blueone]

According to Sambanova website, Tan is still the Chairman.

Hmmm... I must have heard wrong.

 

[KevinK]

I'm only guessing, but I suspect LBT may be buying a team more than a design.

A team that has built a well-optimized, credible rack level product, but needs to field one to two generations of more modern silicon, would be valuable. But worth far less than $5B. Plus LBT has to be careful of the optics for a couple of reasons - his conflict of interest, plus the cash infusions from the US government and NVIDIA.

 

[blueone]

A team that has built a well-optimized, credible rack level product, but needs to field one to two generations of more modern silicon, would be valuable. But worth far less than $5B. Plus LBT has to be careful of the optics for a couple of reasons - his conflict of interest, plus the cash infusions from the US government and NVIDIA.

I'm not sure what to think about their technology, because there's so little technical detail available. Reconfigurable dataflow processing chips are a thing right now. (NextSilicon is claiming similar capabilities, among a couple of others in early stages I'm aware of.) Central to the advertising of both companies is how primitive, slow, and power-hungry CPUs and GPUs. I'm still curious. Skeptical, but curious.

 

[count]

Tan is going to get rich(er) off of Intel no matter what happens haha

 

[KevinK]

I'm not sure what to think about their technology, because there's so little technical detail available. Reconfigurable dataflow processing chips are a thing right now. (NextSilicon is claiming similar capabilities, among a couple of others in early stages I'm aware of.) Central to the advertising of both companies is how primitive, slow, and power-hungry CPUs and GPUs. I'm still curious. Skeptical, but curious.

Yeah,

There probably is promise, but I'm not sure GPU-based AI processors are the "pigs" that some make them out to be, especially if they can get acceptable FP4 results ahead of 3 generation back hardware stuck at FP16 arithmetic. I really have taken to benchmark watching both from Artificial Analysis when they do simplistic chip vs chip comparisons - here's a dated one.

https://twitter.com/x/status/1927896050112811125

And SemiAnalysis when they do rack-level benchmarks with typical data center loading and TCO per GPU and tokens per MW numbers.

InferenceMAX™: Open Source Inference Benchmarking

NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B

newsletter.semianalysis.com

 

[Xebec]

Yeah,

There probably is promise, but I'm not sure GPU-based AI processors are the "pigs" that some make them out to be, especially if they can get acceptable FP4 results ahead of 3 generation back hardware stuck at FP16 arithmetic. I really have taken to benchmark watching both from Artificial Analysis when they do simplistic chip vs chip comparisons - here's a dated one.

https://twitter.com/x/status/1927896050112811125

And SemiAnalysis when they do rack-level benchmarks with typical data center loading and TCO per GPU and tokens per MW numbers.

InferenceMAX™: Open Source Inference Benchmarking

NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B

newsletter.semianalysis.com

Is that Cerebras using an entire wafer to beat a Blackwell GPU?

 

[blueone]

Is that Cerebras using an entire wafer to beat a Blackwell GPU?

In a way, yes, at least if I understand the test configurations, but a Llama 4 "endpoint" (I think) means this is a "single user" test. So, if I'm correct (60% confidence level) it means for a single user the Cerebras CSE-3 is twice as fast, and the multi-user test implies the Cerebras system scales linearly for ten users, while the Blackwell system doesn't. If I'm correct, the test is structured for maximum single endpoint throughput, so the Blackwell is not partitioned. In the multi-endpoint test it seems to be partitioned. The CSE-3, with about 900,000 cores, scales linearly because it has an over-whelming amount of parallelism compared to a single Blackwell. Also, the cost seems to be calculated to focus on the cost to run single endpoint as a fraction of the total system cost. So, yeah, the CSE-3 is much more expensive and overall consumes more power, but fully utilized it costs less per endpoint than a comparable Blackwell system of similar capability would.

These tests results strike me as convoluted too.

If anyone here has a more accurate interpretation of what all of this marketing stuff means, I'd like to be educated.

 

[Daniel Nenni]

Was chairman, but the answer is yes, though it's not a showstopper. The classic example: Steve Jobs, Apple, and NeXT.

Not at all. Intel has an M&A team and lawyers that do the work. The same thing happened at Cadence. A couple of the companies Cadence acquired had Walden Ventures inside. Hopefully the M&A team does their job. I did an acquisition with Cadence and their team was very good. Siemens has the best M&A team I have ever worked with. They do a LOT of deals. I have not done a deal with Intel but I have heard no complaints.

 

[Daniel Nenni]

Is that Cerebras using an entire wafer to beat a Blackwell GPU?

I interviewed an executive from Cerebras a while back and I asked him how many cores they had on the die/wafer. The said they would guarantee x amount of working cores but would not say how many and I asked multiple times. If only half of the cores work why brag about using the whole wafer?