Whatโs ๐ฅ in Enterprise IT/VC #431
๐๐ผ DeepSeek - years compressed into days - the cost ๐ฐ of intelligence ๐ง has dramatically ๐ - the time to build ๐๏ธ is now!
It was hard to go an hour without talking to someone about DeepSeek this past week, and while the drama was off the charts, hereโs where folks are netting out.
The cost for reasoning is going to get dramatically lower in the future.
To be fair, no US company will use the DeepSeek API as all of that data will be used by China. However, hereโs the gift ๐, itโs open source/weight model and under the incredible MIT license which means any one can download, enhance, and add to commercial products without any fees. Since itโs also memory efficient, one can even deploy on the edge ๐คฏ. This 95%+ drop in costs in reasoning for a model which is supposedly good enough is fantastic for all founders building up the AI stack above the foundational model layer. Since itโs open source, weโve already seen Microsoft Azure offer DeepSeek and also Dell.
Why you may ask? Clearly the customers are asking for this option.
Even AWS is getting pressure to offer DeepSeek on Bedrock (Insider)!
More than 20 key clients of Amazon Web Services asked the company to make DeepSeek models available through Amazon's Bedrock AI development tool this weekend, according to an internal document obtained by Business Insider. Toyota, Stripe, Cisco, Yelp, and Workday were among AWS customers asking for this access, with many wanting to test and evaluate DeepSeek's AI capabilities internally. Other companies that made similar requests include Mercado Libre and Kellogg, the document showed. An Amazon spokesperson told BI that Bedrock customers use multiple models to meet their unique needs and the company remains focused on "providing our customers with choice."
And here it is! As an aside, the speed at which these large companies moved has been incredible!
This is unequivocally a massive net positive for all founders and builders assuming youโre not invested in semis or the models themselves. Ok, a bit sensationalist below, but yeah short-term there is some work to be done by the model providers who, of course, will just incorporate any interesting techniques from DeepSeek into newer models.
Hereโs Zuck from this past week reiterating the CapX spend for 2025 (The Verge).
Zuckerberg noted that โthereโs a number of novel things they did weโre still digestingโ and that Meta plans to implement DeepSeekโs โadvancementsโ into Llama. DeepSeek caused a massive sell-off in AI stocks due to fears that models will no longer need as much computing power. Zuckerberg tried to dispel concerns that the billions of dollars heโs spending on GPUs will go to waste: โI continue to think that investing very heavily in CapEx and infra is going to be a strategic advantage over time.โ
In essence, the cost of reasoning and intelligence ๐ง is rapidly decreasing, years compressed into a matter of days, benefiting us all. DeepSeek just forced everyoneโs hands to play ball ๐๐ผ.
As Iโve always said and also shared in my interview with Jesus Rodriguez in
last week, value has been accruing up the stack. For reference, hereโs an excerpt which you can read in full here.Builders now have an open source model which is insanely fast and cheap in which they can easily customize, train on proprietary data, and only pay for the cost of inference. Look at whatโs happening on Hugging Face - something quite powerful.
DeepSeek's advancements in multi-token prediction, mixture of experts architecture, reinforcement learning with minimal human interaction, along with memory optimization writing every number with 8 decimal places instead of 32, will significantly reduce the cost of closed models for everyone.
There will be no need to optimize cost per query with outdated models anymore.
This is the greatest time to be a founder and itโs time to build ๐๏ธ. Jevonโs paradox is in full swing as Francois reposted Satyaโs twist with his own commentary.
Hereโs Andrew Ng chiming in as well.
The byproduct of all this is the hope that we get back to normalized Inception rounds - resharing this 1 minute clip from October as it covers lots of relevant ground.
Hereโs Chamath chiming in to this narrative as well. Letโs get back to โnail it, then scale itโ instead of โscale it, then nail it.
Case in point - Agentuity raised a $4M Inception round - got to love those repeat founders who simply donโt want too much cash to start.!
As value continues to accrue up the stack, prepare for the tidal wave of agents ๐ค coming your way. Which is why Iโm ๐ฅ up about our investment in Agentuity. Think Heroku for Agents +++ . Read ๐งต
As Jeff Haynie states:
Bottom line - thank you DeepSeek, this is the best time to be a founder ever! Anything is possible if you believeโฆ
Hereโs my quote from The Information this week:
Yes, a bit sensationalist, and to clarify, my quote should say:
โFounders now just need a DeepSeek download, a Stripe account, three to four people, and you can create an insanely capital-efficient business.โ
Any how you get my point, and now that DeepSeek account can be an Azure account to get API access or a GCP account with a DeepSeek API which of course will alleviate any privacy and data concerns with using an API of a Chinese-based company.
Bottom line, we are deploying dollars ๐ฐ at boldstart and so ๐ฅ up for the future!
As always, ๐๐ผ for reading and please share with your friends and colleagues.
Scaling Startups
#what YC is looking for - lots of agents!
#the energy just jumps off the page, amazing podcast with my boldstart partner Eliot Durbin with Tim Chen and Ian Livingstone, a founder weโve partnered with for over a decade!
#take that leap now - you donโt have all of the answers, none of us do, esp. with how fast AI is changing everything
#intertia is the strongest force so when you start generating some momentum, important to keep it rolling
#enterprise marketing is changing super fast
Enterprise Tech
#regular readers know Iโve been thinking about second order effects of AI and what happens when hundreds of thousands of agents run amok at enterprises and when AI coding becomes >50% of all code generated. Well, this is why Iโm super excited about Baz coming out of stealth this past week! Itโs founded by Guy Eisenkot and Nimrod Kor out of Israel ๐ฎ๐ฑ. Guy was previously co-founder of Bridgecrew, sold to Palo Alto Networks for 200M, and Nimrod worked closely with him there as software eng manager driving a lot of the development of Checkov, one of the most popular devsecops git repos. Any how, more code equals more code to review and Bazzy the code review agent is here to save the day!
And hereโs our boldstart why we invested post to give you a lens into Inception investing.
#whatโs the breakthrough in Deepseek technically - here are 2 best breakdowns
from Wordgrammer ๐งต
I donโt think thereโs anything magical here. I really think they just made 2 massive cost-cutting innovations, which let them run more experiments, which led them to reverse engineer o1 faster.
#and also โdistillationโ - for those who donโt know what it is (WSJ)
The Chinese companyโs leap into the top ranks of AI makers has sparked heated discussions in Silicon Valley around a process DeepSeek used known as distillation, in which a new system learns from an existing one by asking it hundreds of thousands of questions and analyzing the answers. โItโs sort of like if you got a couple of hours to interview Einstein and you walk out being almost as knowledgeable as him in physics,โ said Ali Ghodsi, chief executive officer of data management company Databricks. The leading AIs from companies like OpenAI and Anthropic essentially teach themselves from the ground up with huge amounts of raw dataโa process that typically takes many months and tens of millions of dollars or more. By drawing on the results of such work, distillation can create a model that is almost as good in a matter of weeks or even days, for substantially less money.
#irrespective of the breakthroughs from DeepSeek, AI is compute hungry! (from founding director OpenAI)
#why more compute will be needed over time
#history rhymes - the comments are fascinating!
#second order effects of agents multiplying - as Iโve always said, we need to rethink lots of infra to optimize for this new world
#Dharmesh crushing it last weekend - yes, we are investing in all of these areas at boldstart and check out MCP.run from port co Dylib.so!
#all aligns with how weโve been thinking about investing in agents
#product decisions from VP Product OpenAI ๐๐ผ - healthy comments/debate
#the power of open source and Deepseek - that didnโt take long ๐คฃ
#this is the year of AI security
Markets
#what me worry says OpenAI
#Eleven Labs raises $180M at $3.3B valuation ๐คฏ just a year after it raised at $1B valuation (TechCrunch)
ElevenLabs, one of the more popular startups working in the field of AI audio, said Thursday that it has raised a Series C round of $180 million, valuing the company at $3.3 billion post-money. a16z and ICONIQ Growth are co-leading the investment. At a time when startups are still finding it challenging to close growth rounds, itโs notable that ElevenLabs raised its Series B round of $80 million, which valued it at $1 billion, just a year ago. ElevenLabs has raised a total of $281 million to date.
#be awareโฆ