Revenue Builders · 30 Jul 2026 · From the week of 27 July
Why Partner Leaders Must Master the AI Landscape to Win Hyperscaler Partnerships with Alan Chhabra
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Alan Chhabra, Executive Vice President of Worldwide Partnerships at Cerebras Systems, returns to the Revenue Builders podcast hosted by John McMahon and John Kaplan. He argues that partner leaders in AI must first map the full stack, from countries, power and chips through hyperscalers, models, software companies and integrators, before building partnerships. He describes how Cerebras got into AWS by pitching a joint architecture alongside Trainium, and how hyperscaler decisions are shaped by politics and their Nvidia relationships. He also covers how he would screen partner leader candidates and why he sees software companies moving faster than traditional enterprises on AI.
For founders
- Test partner leader candidates on whether they can lay out the AI landscape of potential partners and explain why each one matters to your company.
- Treat technical understanding as a non-negotiable screen for any partner leader at an AI company.
- Look for candidates who have opened hard partnerships before, rather than those who have only maintained relationships others opened.
- Frame any hyperscaler offer as a long-term win for them, since their decisions are shaped by Nvidia supply relationships and their own chip programs.
- Ask candidates to show the top-level org chart of your most important partner to test how well they understand who holds power.
For revenue leaders
- Map the full AI stack, from power and chips to models, software and integrators, before trying to crack a hyperscaler.
- Find an angle the hyperscaler's own technology supports, such as pairing its chip with yours, so the deal is a win for them.
- Playing hyperscalers against each other at MongoDB created competitive pressure that drew in Google, then AWS, then Microsoft.
- A partner leader should be willing to personally reach senior executives at key partners and be the one who asks partners for hard things.
- Expect software companies to move faster than traditional enterprises on AI adoption, which is shifting where sales leaders spend their time.
What was said 23, most useful first
A partner leader should map the full AI stack, from countries and power through chips, hyperscalers, models and software, before trying to crack a hyperscaler. Listen
Alan describes a pyramid with countries and land at the base, then power, water and minerals, then chip fabrication (he names TSMC), then chip makers including Nvidia and Cerebras, then data center builders, hyperscalers, neoclouds and colos, then model providers, software companies, system integrators and customers. He says he had to work this out for Cerebras before he could get into a hyperscaler, and recommends that every AI partner leader do the same for their own landscape.
“you need to figure out your own landscape, only then can you go crack into one of the ranks of the pyramid”
Listen to the episode Strategy & market Link to this Report a problem
Cerebras won an AWS deal by pitching a joint architecture that pairs Trainium with Cerebras, so AWS's own chip stays central. Listen
Rather than asking AWS to add a third-party chip, Alan says the architecture splits the work: Trainium handles prefill, the input of tokens, and Cerebras handles decode, the output. He says this leverages AWS's own IP and is competitive on price performance, and that AWS has signed a big deal with Cerebras which is being rolled out this year.
“You really have to come up with the angle that they're going to buy into. That's a win for them long term.”
Listen to the episode Strategy & market Link to this Report a problem
At MongoDB, signing one hyperscaler first created competitive pressure that brought the others into the partnership. Listen
Hyperscalers were offering their own open source database as a service, so MongoDB asked them to sell its enterprise versions instead. Alan says Google signed first, and when AWS was competing with Google it then wanted in, followed by Microsoft, with the pitch that MongoDB's sales teams could back a hyperscaler or compete with it.
“So which hyperscaler wants to get with us first?”
Listen to the episode Strategy & market Link to this Report a problem
By the time Alan left MongoDB, at least half its revenue went through a cloud marketplace and around 90% was touched by partners. Listen
Alan says he built MongoDB's partner program to get other people to sell what MongoDB sold, aimed at helping its sales force. He says the company is now over $2 billion in revenue.
“at least 50% of the revenue that MongoDB had went through a cloud marketplace and I think 90% was touched by partners”
Listen to the episode Strategy & market Link to this Report a problem
Cracking hyperscalers is extremely political, because they depend on Nvidia supply and are building their own chips, so technical merit alone does not decide. Listen
Alan says hyperscalers have deep relationships with Nvidia and do not want to upset that supply chain. When a host notes that in a shortage Nvidia could put a hyperscaler on the slow boat, Alan agrees it is a lot about supply and demand. He also says hyperscalers are building their own AI chips, such as Trainium, Microsoft's Maia and Google's TPUs, so they want to stay relevant to the stack.
“cracking the hyperscalers is extremely political”
Listen to the episode Strategy & market Link to this Report a problem
A CEO can test a partner leader candidate by asking them to lay out the landscape of potential partners and why each matters to the company. Listen
Alan says a CEO of an AI company should ask this question before anything else, and that a candidate who cannot answer has not done their homework. He says anyone can name-drop four people, but the landscape question shows whether they understand the players and where they can be played against each other.
“Ask them the interview question. What's the landscape of all the potential partners we could have and why relevant to us?”
Listen to the episode Hiring & team building Link to this Report a problem
A CEO can ask a partner leader candidate to show the top-level org chart of the most important partner to test their grasp of who holds power. Listen
Alan says he uses this exercise himself, and when he joined Cerebras he showed Andrew Feldman the power structure inside AWS. He describes it as several competing kingdoms reporting up to different leaders, and says the candidate should show the top level, not every employee.
“give me the org chart, not like all 20,000 people, but at the top level, show me the org chart of the number one partner you're going to work with at our company”
Listen to the episode Hiring & team building Link to this Report a problem
For partner leaders at AI companies, candidates who do not understand the technology should be disqualified right away. Listen
Alan says most AI companies are engineering-heavy, so a partner leader without technical grounding will be seen as the stepchild. He says he used to believe technical skills could be taught to smart people, but no longer does, and he does not require two engineering degrees, only real technical fluency. He says relevant backgrounds include understanding piping, cooling and water.
“if you are a candidate that doesn't understand technology in this landscape”
Listen to the episode Hiring & team building Link to this Report a problem
In AI, the infrastructure layer may be the center of the market for partners, not the applications built on top of it. Listen
Alan contrasts the database era, when open source databases were the center, with the gold rush, where the people selling pans and shovels made money either way. He says that about two years ago he concluded Nvidia was the center, and that anyone who could create an alternative architecture would be viable, which led him to Cerebras.
“And I'm like, you know what? That's the center.”
Listen to the episode Strategy & market Link to this Report a problem
Training AI models alone earns little return until the model is used for inference. Listen
Alan says the early AI market made money selling GPU training, but training alone has little ROI because nobody uses the model until it serves answers. He says demand now comes in text, pictures, PDFs and voice, and needs millisecond responses, which is why he pitches Cerebras on fast inference.
“Training alone, there's little ROI because no one's actually going to use the model until you get inference.”
Listen to the episode Strategy & market Link to this Report a problem
Cerebras's chip is in many cases ten times faster than GPUs for AI inference, which Alan uses as the core pitch to hyperscalers. Listen
Alan says the chip's architecture makes it in many cases 10 times faster at AI inference than anything on GPUs. He argues that a hyperscaler competing for AI customers would want this performance, and presents it as the company's own pitch.
“it is in many cases 10 times faster the AI on this than on anything on GPUs”
MongoDB's partnership with Alibaba in China gave it leverage with the hyperscalers, because rivals did not want MongoDB to partner with the Chinese before they had a deal. Listen
Alan says MongoDB cracked Alibaba in China and that the last thing AWS or Azure wanted was MongoDB having a strong Chinese partnership before they had one of their own. He presents this as one way of building political capacity across competing relationships.
“And the last thing AWS or Azure wanted was MongoDB to have a great partnership with the Chinese before we had a good one with them too.”
Listen to the episode Strategy & market Link to this Report a problem
A partner leader should be willing to personally reach senior executives at key partners, including by texting a cell phone number they found. Listen
Alan says he prospects senior partner executives himself rather than delegating it, including through LinkedIn. He says he recently found a leader's cell phone number, without saying how, and texted them to set up a call with his CEO. He says there is a skill to it, that AI makes it a little easier, and that a leader has to want to do it and be able to teach others.
“I do it through LinkedIn. Yesterday, I found a leader's cell phone number.”
Listen to the episode Pipeline & demand generation Link to this Report a problem
The top partner leader in a company may need to be the one who asks partners for things no one else will, even if it makes them the bad guy. Listen
Alan says it is okay to pin partners against each other as long as it is done fairly, to ask for more, and to stand up to protect the company. He says the top partner leader should be respected but should not be the one who does the fireside chats and keynotes.
“It's okay to ask for more. It's okay to be the bad guy.”
Listen to the episode Leadership & culture Link to this Report a problem
Look for partner leaders who have opened hard partnerships before, since opening is a grind that wears people down. Listen
Alan agrees with a host that many partner people only manage relationships someone else opened. He says most AI partnerships are not yet open, but finding people willing to repeat the grind of opening tough partnerships is hard. He says he finds opening a partnership exciting and gets bored when one is sailing smoothly. He describes his own path from services and sales into partners after Dave told him to run partners.
“I personally find it exciting to open something and I get a little bit bored when it's, you know, kind of sailing smoothly.”
Listen to the episode Hiring & team building Link to this Report a problem
Alan expects governments to keep pouring money into in-country model providers, which he sees as a competitive advantage for each country. Listen
He lists Cohere in Canada, Mistral in France and Alibaba's models in China as examples of providers that countries back. He compares AI output to the size of a military or oil reserves, and says partners selling into these countries need to understand that the choice is political, not only about the best technology.
“governments are going to pour more money into those because they see that as a competitive advantage for the future”
Listen to the episode Strategy & market Link to this Report a problem
Neoclouds could win sovereign AI deals in countries such as Switzerland because they can move faster than other providers. Listen
Alan says neoclouds like the sovereignty concept because they are the newest and most agile. He gives a hypothetical in which a country with a government-funded sovereign AI initiative sees a neocloud build a data center faster than anyone else, which he says could lead to very large deals.
“a neocloud could swoop in and say, I can win that business”
Listen to the episode Strategy & market Link to this Report a problem
Software companies are embracing AI and accepting IP risk, because waiting is an existential risk, while traditional enterprises are moving slowly to protect data and IP. Listen
Alan says that for a software company he goes on the side of innovation every time, arguing that the company could be gone if it waits. For an enterprise like a bank he would err toward protecting sovereign data and getting it right before going all in.
“Yeah, software companies are willing to embrace AI, willing to take the risk, either use it or get eaten by it.”
Listen to the episode Strategy & market Link to this Report a problem
Most AI-space sales leaders now sell more to software companies than to traditional enterprises, which is the reverse of ten years ago. Listen
Alan says he holds this view from his conversations with sales leaders, including friends such as Alex Verrill. He attributes the shift to software companies embracing AI while traditional enterprises remain cautious about data and IP protection.
“Most sales leaders today, probably at least in the AI space, are selling to software companies more than they used to versus traditional enterprises because of what I said.”
Listen to the episode Strategy & market Link to this Report a problem
Software companies can let anyone install their product through AI and still sell the product itself, while dropping implementation services. Listen
Alan says MongoDB lets anyone install its database with AI, which he calls a smart use of AI. He says the company still has the database to sell but can no longer sell implementation services for it.
“But you still have the database that you can sell them. You just can't sell them any implementation services.”
System integrators have moved higher in the AI stack and are becoming less relevant, because AI is replacing the application-building work they do. Listen
Alan says he used to place system integrators low in the stack beside hyperscalers, but they now sit above models and software because they help customers build applications with AI. He notes a conflict, since AI is replacing that work.
“But in many cases, they're also becoming less relevant because AI is replacing them.”
Listen to the episode Strategy & market Link to this Report a problem
Alan expects the hyperscalers to win the enterprise AI market, with banks and other regulated enterprises running AI on them rather than on neoclouds or government-built clouds. Listen
He says neoclouds focus more on ISVs, while hyperscalers have rigor on data security and software. When a host suggests regulation may mandate this for banking, Alan agrees. He compares it to banks and insurers moving last to the cloud before making giant deals.
“I still think the hyperscalers win the enterprise.”
Listen to the episode Strategy & market Link to this Report a problem
Alan expects AI data centers to shift toward liquid or water cooling, so existing facilities may need retrofitting. Listen
He says air conditioning will not be how most AI computing is cooled in the future, and that data centers must be built or retrofitted to handle liquid or water cooling. He points to the piping work this requires.
“Air conditioning is not the way most AI computing will be cooled in the future. It'll be through liquid or water.”