Revenue Builders · 30 Apr 2026 · From the week of 27 April
How AI Is Rewriting the Sales Playbook and Raising the Bar on Human Performance with Alex Varel
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Hosts John McMahon and John Kaplan interview Alex Varel, EVP of Worldwide Sales at Cerebras Systems, about selling AI infrastructure where inference speed and memory bandwidth are the main constraints. Alex explains the shift from AI training to inference, why latency targets of around 500 milliseconds matter for conversational products, and how that connects to customer value. He describes how sellers should qualify, orchestrate cross-functional buying committees, use agents to reduce non-selling admin, and keep the ICP under constant review. His central argument is that core sales disciplines persist, but the playbook has to adapt quickly as technology and buyers change.
For founders
- Alex argues that once models reach reasoning parity, the remaining differentiator is the speed and scale of the inference engine.
- Alex cites a conversational latency target of about 500 milliseconds or less, and says a two second response is too slow for conversation.
- Alex's team treats the ICP as a live input, updating target lists from capital flow and funding signals rather than once a year.
- In his technical sale, the AE has to orchestrate infrastructure owners, product leaders and engineering to turn a technical win into a business win.
- Alex points to premium speed tiers already appearing, such as Anthropic's fast option and OpenAI's Codex Spark powered by Cerebras, and suggests customers on flat subscriptions could open a premium tier that pays for the speed.
For revenue leaders
- Ask why a prospect wants to know before answering a question such as which models you serve, because the reason shapes the roadmap discussion.
- Host John McMahon says sellers spend only 25 to 30 percent of their time selling face to face and believes AI can help move that toward 50 to 75 percent, a point Alex agrees with.
- Connect agents to the system of record so they can run ongoing research and surface patterns and risks, and keep the tool stack lean.
- Translate technical metrics into customer satisfaction and revenue impact by asking what the metric changes for the business.
- Hire sellers whose skills transfer across technologies, which Alex did by recruiting from Snowflake and EMC for a novel market.
What was said 23, most useful first
Once models reach reasoning parity, the remaining differentiator is the speed and scale of the engine running them. Listen
Alex argues that AI models are converging and many applications overlap, so he sees speed and scale of the engine, and the sophistication of workloads it can support, as what is left to differentiate. He makes this point from his role at Cerebras Systems, and says he does not mean to be controversial when he suggests that hardware sellers were once dismissed as box pushers.
“To me, the real differentiator left is the speed and scale of the engine and the sophistication of the workloads it can support.”
Listen to the episode Strategy & market Link to this Report a problem
For conversational AI, Alex puts the latency target at about 500 milliseconds or less. Listen
He describes a pharmacy's conversational bot with a two second response target, which he calls way too slow for conversation, and says conversational use needs roughly 500 milliseconds or less. He says Cerebras powers instant AI for conversational, agentic coding and deep search use cases. He describes asking a prospect what changes if responses arrive in 200 milliseconds against a 500 millisecond target.
“You need to be like 500 milliseconds or less, right?”
Start value conversations from the technical metric and ask what it changes for customer satisfaction and revenue. Listen
Alex says his strength is business acumen, so when he has a latency advantage he asks prospects where it would affect customer satisfaction, revenue and other outcomes. A host then highlights the story, saying technical requirements create business outcomes and that people moving into AI native companies can get stuck at the technical discussion.
“Can you help me understand where that might have an impact on? your customer satisfaction, your revenue, you know, on and on and on.”
Listen to the episode Positioning & marketing Link to this Report a problem
Most sellers spend only 25 to 30 percent of their time selling face to face, with the rest on admin and research. Listen
McMahon says the remaining 75 to 80 percent goes to admin, updating the tech stack, researching customers and use cases, and coordinating calendars and information. He believes AI can at least halve that non-selling work, which could lift selling time to 50 to 75 percent. Alex agrees.
“I'm spending 80 % of my time doing that crap that AI can help me at least cut in half, maybe even more.”
Listen to the episode Sales process & deals Link to this Report a problem
A RevOps lead built five or six agents in one weekend for voice of customer, forecasting and solution architecture monitoring. Listen
Alex describes Jimmy Lee, his RevOps lead, who trained five or six agents over Easter weekend. One agent moves customer feedback from JIRA tickets into product feedback, another supports forecasting, and another monitors solution architecture work, with six more in progress. Alex says this kind of setup should multiply productivity.
“He's got one for voice of the customers to product feedback going through JIRA tickets and things like that.”
When a prospect asks a question such as which models you serve, ask why they want to know before answering. Listen
Alex gives the example of a customer asking which models Cerebras serves, which he says may reflect concern about the model roadmap, since models change quickly. He says he likes to hire sellers who qualify and listen for intent. A host adds that the best sellers are skeptical qualifiers who weigh whether to answer fully, partially or not at all, and Alex agrees.
“Can you help me understand? Why why did you ask?”
Listen to the episode Sales process & deals Link to this Report a problem
Treat the ICP as a live input updated from capital flow and funding signals rather than an annual exercise. Listen
Alex says his team tracks the flow of capital and raises, pulls data from multiple sources, synthesizes it, and constantly updates top target lists. Scoring and market signals move accounts into tier one, which takes most seller time, and so change how sellers prioritize their pipeline cadence.
“We're tracking the flow of capital. We're tracking raises.”
Listen to the episode Strategy & market Link to this Report a problem
The AI market has shifted from model training to inference, which moves competitive advantage toward speed. Listen
Alex says massive resources were spent training the models people use now, but requests are now more frequent and more complex. He argues the real engine behind the user experience is the speed of inference, so that is where competitive advantage now sits.
“the world has shifted from AI training to AI inference”
Inference is limited by memory bandwidth, because each generated token requires moving the model's weights again. Listen
Alex gives the example of a 70 billion parameter model, which he says moves 140 gigabytes of data from compute to memory to generate one word, with the weights moved again for the next word. He says Cerebras's architecture, which keeps memory and compute on one very large chip, addresses this memory wall bottleneck.
“And the whole bottleneck with AI inference is memory bandwidth”
Users expect AI to respond at human conversational speed, so slow AI responses have no market. Listen
Alex compares AI to search and says there is no market for slow search. He expects agentic AI and reasoning to be held to the same human clock speed expectation, and says no number of GPUs stitched together can match that responsiveness.
“What is the market for slow search? There isn't one.”
Agents can take on RevOps work, so the RevOps team Alex expected to need is smaller than planned. Listen
Alex says the RevOps team he thought he would need 14 months ago is not what he needs now, because automation and agentic workflows give back significant time. He says he has fewer software tools and RevOps staff than he traditionally would, and that agents must be company-approved and connected to the system of record.
“I don't have nearly the software stacked. or the Rev Ops staff that I traditionally would have because we've found automation and agentic workflows that are really helpful.”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
Connect agents to systems of record so they run ongoing research and surface patterns and risks for sellers. Listen
Alex says he runs a traditional CRM and content management system, has some automation in outreach, and otherwise relies on agents his team is training. These agents have access to company systems, pull from systems of record, do ongoing research and analysis, and surface patterns and risks, and he says his stack is shorter and has fewer people behind it than any he has had.
“We are training agents, the ones that have access connected to our systems to pull from our systems of records and do some ongoing research and analysis and surface patterns and risks to us.”
Before buying AI tools, check that each use case maps to a specific high-value sales workflow that makes sellers more productive. Listen
The host says you should not buy technology just to buy it or to have relationships. He says each use case should be matched to a high-value sales workflow that makes sellers more productive, starting with areas such as qualification, and that tools should reach curious sellers seamlessly and with little pain.
“You're not buying technology just to buy it or to have relationships.”
The AE must orchestrate a cross-functional buying committee, since technical wins alone do not close the business case. Listen
Alex says that after the technical team wins the math, the business win still has to be built with infrastructure owners, product leaders, engineering and other technical experts, each of whom can win logic battles. He says the AE must listen, find where the value is and command a premium for the fastest tokens.
“But the AE has got to be an extreme orchestrator.”
Listen to the episode Sales process & deals Link to this Report a problem
Strong technical knowledge is not enough; top sellers also need interpersonal intelligence and curiosity. Listen
Alex draws on Howard Gardner's framework that human potential is not a single IQ score, and argues that AEs need several kinds of intelligence. He says a founder-led or purely technical approach misses the curiosity needed to ask why a customer is asking a question.
“People can look that up that human potential isn't just like one IQ. score, right?”
Listen to the episode Hiring & team building Link to this Report a problem
Recruit sellers whose skills transfer across technologies, drawing from Snowflake and EMC for a novel market. Listen
Alex says he recruited from Snowflake for elite sellers who had worked on a novel architecture, and likes EMC sales culture. He says what he needs are the transferable intangibles, because the new market is novel territory for everyone, including the sellers.
“I went and pillaged Snowflake. I mean, what an incredible company.”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
Five constants persist across markets: generating pipeline, qualifying it, messaging, managing ICP and territory, and forecasting. Listen
Alex lists these as the work sellers must do regardless of technology. He says the customer and how they buy change, and that AEs must learn to work with a machine learning solution architect in much the same way they work with solution engineers on other technologies.
“We're going to have to go out and attack the market and generate a pipeline. We're going to have to have the courageous ability to qualify the pipeline.”
Listen to the episode Strategy & market Link to this Report a problem
Well-funded AI native companies that emerged in the last six to eighteen months are now priority targets alongside the Fortune 500. Listen
Alex says his instinct was to attack Fortune 500 and Forbes 2000 accounts, which he describes as scared but slower moving. He says well-capitalized AI native companies have disrupted the market, and his team builds pipeline behind the ones it believes will do well over the long term.
“it's hard to tell like and who's going to do great things and create new categories and new economies.”
Listen to the episode Strategy & market Link to this Report a problem
Alex suggests AI companies with flat subscriptions could add premium speed tiers, as some providers already have. Listen
Alex describes a customer with a flat, one-size-fits-all subscription and asks whether it could open a premium tier for speed. He notes this is happening elsewhere, citing Anthropic's fast option and OpenAI's Codex Spark, which runs on Cerebras. The customer said it had never considered premium tiers but agreed it was possible, and Alex said he imagined such a tier would more than pay for the speed.
“And I imagine it would pay for what you and I are talking about in droves.”
Listen to the episode Pricing & packaging Link to this Report a problem
Buyers will value human seller traits as AI-written email becomes generic. Listen
Alex says email has been taken over by AI slop, and predicts that buyers will crave human fallibility, custom-made emails and face-to-face meetings. He argues sellers should keep their interpersonal and communication skills while also learning to work with agents.
“The email is dead. It's just been taken over by AI Slop.”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
Sellers should show their productivity with agents working on their behalf, not only their own work. Listen
Alex says he would encourage his daughter to present agents that act as multipliers of herself in an interview, similar to an artist showing a portfolio. He says sellers without hands-on experience augmenting their productivity or supervising agents will be less capable.
“I want her to come and show the agents that are multipliers of herself, extensions of herself”
Listen to the episode Hiring & team building Link to this Report a problem
Token budgets per employee may become a labor consideration as AI use spreads. Listen
Alex says there might be a token budget per employee in the future, and notes there kind of is one now. He calls this a labor shift, not just a software shift, and suggests leaders understand prompt engineering and tokens and build a self-learning enablement path.
“There might be a token budget per employee in the future, right?”
Attach selling to the biggest business issue facing customers and influence their decision criteria with your differentiation. Listen
Kaplan says he has been teaching this for decades and it matters more now at Force Management. Alex agrees it remains true in the new AI environment, and says it holds regardless of how the technology changes.
“Attach yourself to the biggest business issue facing your customers and influence their decision criteria with your differentiation.”
Listen to the episode Sales process & deals Link to this Report a problem