Learning Library

← Back to Library

AI Roadmap 2026: Compliance Opportunities

Key Points

  • 2026 AI planning now requires anticipating five key trend drivers, starting with tightening regulatory enforcement worldwide.
  • The EU AI Act will roll out enforcement from August 2025 to full compliance by August 2027, while California and over 45 U.S. states are passing AI bills that impose transparency, safety, and hefty penalty requirements.
  • Rather than just a cost, compliance creates a new market opportunity, demanding robust measurement practices such as bias and performance testing, model cards, evaluation packs, and forensic audit capabilities.
  • Smaller firms must scale compliance investments appropriately, while larger organizations and emerging vendors can capitalize by offering specialized compliance infrastructure and services.

Sections

Full Transcript

# AI Roadmap 2026: Compliance Opportunities **Source:** [https://www.youtube.com/watch?v=x_fsaOnqbeo](https://www.youtube.com/watch?v=x_fsaOnqbeo) **Duration:** 00:28:49 ## Summary - 2026 AI planning now requires anticipating five key trend drivers, starting with tightening regulatory enforcement worldwide. - The EU AI Act will roll out enforcement from August 2025 to full compliance by August 2027, while California and over 45 U.S. states are passing AI bills that impose transparency, safety, and hefty penalty requirements. - Rather than just a cost, compliance creates a new market opportunity, demanding robust measurement practices such as bias and performance testing, model cards, evaluation packs, and forensic audit capabilities. - Smaller firms must scale compliance investments appropriately, while larger organizations and emerging vendors can capitalize by offering specialized compliance infrastructure and services. ## Sections - [00:00:00](https://www.youtube.com/watch?v=x_fsaOnqbeo&t=0s) **2026 AI Roadmap: Regulatory Driver** - The speaker urges businesses to start planning for 2026 by outlining five key AI trend drivers, beginning with the first—intensifying AI regulations like the EU AI Act and California's bill—that create new compliance market opportunities. - [00:04:02](https://www.youtube.com/watch?v=x_fsaOnqbeo&t=242s) **Economic Scrutiny Drives Outcome Pricing** - Growing macro‑economic pressure and the rise of AI agents will force vendors to prove ROI through outcome‑based pricing across sales, product, and engineering, or risk losing market share by 2026. - [00:08:27](https://www.youtube.com/watch?v=x_fsaOnqbeo&t=507s) **Incumbent vs Disruptor ROI Strategies** - The speaker contrasts how established firms can capitalize on preserved data and context to ensure smoother productivity gains, whereas disruptors must prove a dramatically higher, easy‑to‑switch ROI—often amplified by AI agents tied to high‑value workflows. - [00:11:34](https://www.youtube.com/watch?v=x_fsaOnqbeo&t=694s) **Hybrid On‑Device AI Emergence** - The speaker predicts that by 2026 Nvidia‑Intel investments will enable on‑device hybrid AI architectures, prompting split cloud‑local workloads, privacy‑centric LLMs, and a surge of new hardware startups. - [00:15:54](https://www.youtube.com/watch?v=x_fsaOnqbeo&t=954s) **Premium AI as Competitive Edge** - The speaker explains that purchasing OpenAI’s premium, high‑cost tiers grants individuals and companies vastly superior, near‑instant task automation—effectively letting them “be in two places at once”—creating a nonlinear competitive advantage over those on free or lower‑priced plans, and raising issues of access control, UI, and trust. - [00:19:09](https://www.youtube.com/watch?v=x_fsaOnqbeo&t=1149s) **Premium vs Commodity AI Strategies** - The speaker outlines how to align tooling and pricing with customer willingness to pay, advocating cheap, fun “nano‑banana” AI for mass‑market adoption while reserving deep, high‑cost capabilities for premium, enterprise users. - [00:23:05](https://www.youtube.com/watch?v=x_fsaOnqbeo&t=1385s) **Vertical Expertise Over Horizontal Tools** - The speaker warns that broad, generic solutions are increasingly at risk, emphasizing that deep vertical knowledge—such as specialized memory requirements and regulatory navigation—is the defensible advantage, urging companies to focus on a single vertical and highlighting related talent shortages. - [00:27:06](https://www.youtube.com/watch?v=x_fsaOnqbeo&t=1626s) **Anthropic, Microsoft, Google AI Strategies** - The speaker explains Anthropic’s R&D‑driven high‑end compute and Claude‑code tools for enterprise work, notes Microsoft’s concerns about primitive‑level tooling, and warns that Google must monetize Gemini to defend its search revenue as ad spend shifts toward AI chat platforms. ## Full Transcript
0:00is time to think about 2026. If you're 0:02building in the roadmap business in 0:04leadership, everywhere I look across the 0:07business world, people are starting 0:09their 2026 planning right now, which 0:12means we need a view down the road of 0:14the AI trends that are going to matter 0:16the most. I've put together five drivers 0:20for AI trends that I think you need to 0:22be aware of. I'm going to pull them out 0:23into specific implications as we go 0:25through this video and then at the end 0:27I'm going to look through crosscutting 0:28implications that cut across all of 0:30those drivers. Once you get done with 0:32this briefing, you are going to be well 0:34equipped to drive an AI fluent planning 0:37process. Whether you're a product person 0:39building road maps in the seauite 0:41looking at how to provision your team 0:42for next year, buying from a vendor, or 0:45even as an entrepreneur or builder, 0:47figuring out what you want to build or 0:48pivot toward next year. Let's get into 0:50it. Driver number one, regulatory 0:53enforcement is getting teeth and that 0:56creates a compliance infrastructure 0:58market that you can benefit from. 1:00Governance is rapidly becoming a launch 1:02requirement. The EU AI Act enforcement 1:04started in August of 2025 for new GPAI 1:08systems and high-risk systems kick in in 1:11August 2026 with full compliance in 1:13August 2027. So, this is phasing in over 1:1624 months, but you need to start 1:17thinking of it now if you have 1:18operations in Europe. Meanwhile, 1:21California is working on an AI bill that 1:24has implications for everyone because 1:26California's AI bill is going to cover 1:28modelmakers with new transparency and 1:30safety requirements. That happens likely 1:32in January of 2026. Uh there are 45 US 1:36states currently with 550 plus AI bills 1:40somewhere in the legislature. Penalties 1:43are going up too. So these are not bills 1:45that don't have teeth. Penalties are up 1:47to 6% of global revenue in the EU and 1:50average compliance failures cost $9.2 1:53million. These are these are issues with 1:56real teeth. Now, there are ways that you 1:59can derisk, but we need to plan for them 2:01in advance. And so, I want to move from 2:02the sort of scary compliance picture to 2:04what I called out at the top, which is 2:06that compliance is really an 2:07opportunity. First, make sure that 2:09you're measuring correctly. Don't just 2:11measure adoption by loginins. Don't just 2:13measure vibes. Don't just measure 2:15customer satisfaction. You want to be 2:17looking at bias testing. You want to be 2:19looking at performance testing. You want 2:20to have model cards. You want to have 2:22eval packs that might be required to 2:24show the edges and red lines of your 2:26capabilities and where they break down. 2:28You want to have some kind of forensic 2:30or audit capability in place for 2:32incident investigation. Now, of course, 2:34if you are a larger company, this is a 2:36harder and harder requirement. If you 2:38are a smaller company, this is something 2:40that you have to invest in in a way 2:42that's commensurate with your scale. 2:43This is where the builder's opportunity 2:45comes in. There are suddenly going to be 2:48a lot of companies that have questions 2:51about compliance and governance with 2:53their particular model installations. 2:56And we already see vendors out there 2:58pedalling eval packs to help these busy 3:02companies get from I don't know what my 3:05AI is doing to I can show and measure 3:07it. That market is going to grow. I 3:09don't think many of the current vendors 3:11are well positioned because they tend to 3:13be one-sizefits-all. Looking at the wide 3:16forest of of the regulatory environment 3:18with many different kinds of mini 3:20environments across different US states 3:22and across the EU, we will need vendors, 3:25builders, tools that offer very fine 3:28grained eval and compliance 3:30capabilities. This is an a tremendous 3:34opportunity. If you're a builder, 3:35vertical expertise is going to be 3:37defensible. You are going to be able to 3:39build in legal, in healthcare, in 3:41finance and other high regulatory 3:43environments and it will be difficult to 3:45unseat you if you have the reputation of 3:48being a highrusted provider that helps 3:51with compliance in a way that really 3:53lifts the load for IT departments. So 3:56that's the first trend compliance. 3:58Driver number two is economic scrutiny 4:02driven by the macro environment. We are 4:04going to see more and more of a move 4:07toward outcomebased pricing versus seat 4:10licenses. Especially as we move toward 4:13an agentic workforce, you already see 4:15model makers leaning in on AI agents 4:19doing this, that, and the other thing. 4:20Proactive agents, agents that take hours 4:22or 30 minutes or 40 minutes to do their 4:24tasks, agents that can do work end to 4:26end. That is going to put long-term 4:29pressure on pricing plans to focus more 4:32on outcomes. We already see that with 4:34the way tools like Finn and intercom do 4:37outcomebased pricing. I'm suggesting 4:40that we are going to see CFO pressure 4:42for ROI proof push more vendors toward 4:44outcome based pricing beyond just 4:47obvious CS use cases. So we'll see it in 4:49sales, we'll see it in product. I think 4:51we'll start to see it in engineering as 4:53well eventually. If you aren't able to 4:57show your ROI telemetry, if you aren't 5:00able to show a proof of value in the 5:04work that you're selling from an AI 5:06perspective, you are going to have 5:08trouble selling in 2026 because the 5:10competition is going to get fierce. One 5:12of the crosscutting trends here is that 5:14the companies that were funded in 2025 5:17are going to be hitting the ground in 5:192026 and they are going to be fiercely 5:22competing for every sales dollar. And so 5:24you want to be in a position where you 5:27can argue that your product is not only 5:30compliant driver one, but driver two is 5:33able to show 5:36agentic aligned measurable outcomes. If 5:39you're deploying an agent in production, 5:41if you are selling something that has an 5:43agentic element, if you are buying 5:45something that has agents included, you 5:48want to make it clear why that extra 5:50work, that extra token burn, that extra 5:52expense is worth it. And the only way I 5:56can think to make that really math math, 5:59you know, to make 1 plus 1 equal 3 is to 6:04show that the agent can solve problems 6:06that traditional LLM tooling cannot. 6:08That is what we are starting to see 6:11through chat GPT's launch of for example 6:14GDP val which is the new benchmark for 6:17showing that Agentic LLMs can solve real 6:20world tasks in as good a manner or 6:24almost as good a manner as experts in 6:26that field. We will see more progress in 6:28that direction and as we do there will 6:31be an expectation from model makers that 6:33they can charge commensurate to the 6:35value they're providing. All of this 6:38means that multimodel resilience is 6:41going to be essential for cost 6:42management. This was already the case, 6:45but it's going to be especially the case 6:46when you have agents driving variable 6:48price outcomes. You need to be in a 6:50place from an architectural perspective 6:52where you can trade models in and out 6:55very very easily and you do not have to 6:58go through an expensive rearchitecture. 7:00That's also true if you're building 7:01products off AI to sell. You don't want 7:03to be in a position where you're locked 7:04in. You want to be in a position where 7:06you can switch stuff out at the command 7:07line very very easily. Driver number 7:10three, you need to be aware of the 7:13competitive velocity from AI native 7:16entrance. So remember how I said earlier 7:18in this briefing, the dollars that were 7:21funded in 2025 are going to come and hit 7:23the revenue line in 2026. They are going 7:26to hit the competitive landscape in 2026 7:28as well. You should expect a forest of 7:32new AI native competitors to spring up 7:35all around you and launch their products 7:37not just as demos but as fullyfledged 7:40products in the first half of 2026. It 7:44will feel like you are swimming in a red 7:47ocean of competition. If you thought you 7:49were competitive already, it's going to 7:51get worse. You are going to see AI 7:54native entrance claim that they can 7:56persist memory and context, claim that 7:59they can deliver 10x productivity gains, 8:02claim that they can deliver training 8:03advantages. And you know what's 8:05interesting? You can also claim that as 8:07a more established incumbent in the 8:09space if you're thoughtful and build 8:11now. So for example, stateless queries 8:14is something that makes more sense. So 8:16stateless queries are when you ask a 8:18question, you have to load all that 8:19context into the chat. That is actually 8:21something that you're more likely to do 8:23with a new entrant than an incumbent. 8:25And so even if incumbents start to come 8:27with like their big memory packets and 8:28this and that, you can pull a page from 8:30Notion's playbook this week and argue 8:33that your AI solution reinforces the 8:37data that people have already trusted 8:38you with, which is exactly the play 8:40Notion used. And that is part of how 8:42they're making a context play in a 8:45competitive market. And context drives 8:47productivity. And so if we go back to 8:49driver number two and the idea of ROI, 8:52one of the things that established 8:53incumbents can emphasize is that by 8:56preserving data and context in an 8:58existing stack, you can ultimately get a 9:01smoother path to real productivity in 9:04the workforce. Now, if you are on the 9:06other side of the table and you are a 9:07disruptor, you are an entrepreneur, you 9:09are a builder, the key here is getting a 9:13disproportionate gain over the existing 9:15income. You have to be able to show that 9:19you are worth switching to like easily 9:22hands down the easiest solution to 9:24switch to. The bar is very high and 9:25there will be fierce competition from 9:27other disruptors as well. Simple value 9:29propositions that offer tangible 9:31multiples on current ROI are going to 9:34break through the noise and that is the 9:35key for disruptors this coming year. Now 9:38if you are looking ahead this is also 9:40where the agents tie in. So agents are 9:43going to be able to drive a lot of that 9:4610x value for disruptors and they are 9:49going to be tied to golden workflows 9:52that customers will pay for. And so if 9:54you're an incumbent, invest in agents 9:57across the workflows that you know your 10:00customers value the most now. Because if 10:03you don't, you run the risk of a 10:05disruptor coming along and offering that 10:0810x multiple on your current ROI for 10:10customers by building an agent for that 10:13quote unquote golden workflow that they 10:15care about the most. Make sure you know 10:17what those workflows are. Make sure you 10:19map them and make sure you think in 10:20terms of building agentic solutions that 10:23are relevant in your area to stay ahead. 10:26The last thing I will call out is that 10:28the AI native velocity gap is real and I 10:31don't see a path to it closing. 10:33Traditional companies are not moving as 10:36quickly as AI native companies and it's 10:38an inherent risk. AI native startups are 10:41hitting a million dollars in ARR in just 10:436 to 12 months versus 18 to 24 for 10:46traditional SAS. They operate with 50 to 10:4980% fewer employees and they're 10:50iterating 10x faster. It's an inherent 10:53risk factor. And the only inoculation, 10:55the only de-risking that I see as a 10:57possibility is that you are going to be 11:01able to anoint, bless, build an 11:04entrepreneurial team that can act like 11:07an AI native startup inside your 11:10business. Otherwise, you're just going 11:12to move slower and you're going to have 11:14to depend on your data and your context 11:16and your existing distribution 11:18relationships as a moat. Let's get to 11:21driver number four. The maturing 11:23technical market for AI is enabling 11:25production systems that were not 11:27possible even in the first half of 2025. 11:32So there are going to be specific 11:34reasons why this statement is even more 11:36true in 2026 and why that should inform 11:39your road mapping your building plans 11:41and the biggest reason is that 11:43investment that Nvidia made in Intel. We 11:45are moving to a world where we will have 11:48ondevice hybrid architectures. So the 11:51current state really is cloud first for 11:54most consumers and also for most 11:56businesses. Businesses clearly see 11:58liability there. I have lost track of 12:01the number of times I have had 12:03businesses ask me about cloud-based 12:05risk. There is appetite and interest in 12:08local LLMs, but there hasn't been the 12:10compute to support it. Now there will 12:13be. You're going to see NPU laptops, 12:16phones with low latency, high privacy, 12:19and there will be a cost pressure to 12:21bring them down. You're going to see 12:22workloads split likely between the local 12:24and the cloud. You'll have faster UX, 12:27somewhat dumber models on local, and you 12:29will have smarter inference and hardened 12:32pipelines on cloud. One of the things 12:34that I expect to see is a new generation 12:37of builds and startups associated with 12:39the availability of local chips that can 12:42power local LLM experiences. You know 12:44how we have clearly right now that adds 12:46that transparent layer and the LLM looks 12:49across your system. I expect winning 12:52builds that are entirely private onrem 12:55local LLMs that also look across your 12:58system and are always on. They're going 13:00to get to mobile and they're going to 13:01get to the laptop. Now this generates a 13:04lot of downstream opportunities if 13:06you're building. Think about it. You 13:08don't have to assume that you have to 13:10compute cloud inference into all of your 13:13costs. You can think about ways in which 13:17second half of 26 early 27 you are going 13:20to start to be able to make experiences 13:22available that give users choices 13:24between local LLM compute and cloud 13:27compute. you're also going to start to 13:30think about more agentic experiences 13:33cross-pollinating with that local LLM 13:35experience. And so you're going to have 13:37these moments where you have a 13:41human-driven experience in your head as 13:43you start to plan for 2026 and the 13:45customer experience you want. But the 13:48customer is going to have a local LLM. 13:50and that local LLM may act as an agent 13:53for them, may purchase for them, may 13:56shop for them, may develop proactive 13:58recommendations for them. We're already 14:00seeing the tip of the iceberg on that 14:03with products like Chad GPT's Pulse, 14:06which just launched this week. That is 14:08very transparently an ad surface. It is 14:12a surface where you want to start to 14:14drive top offunnel consideration. Now, 14:17couple that with agentic capabilities, 14:19and you're going to start to see agents 14:21reading ads, agents looking at content, 14:23agents proactively taking action, and 14:26those agents may be entirely private and 14:28on the local machine. Lots of 14:31interesting stuff going on there. If you 14:33are thinking about what to build in that 14:35world, build assuming that compute in 14:38the cloud becomes something that is 14:42optional for early adopters in 2026 and 14:45becomes widely spread as an optional 14:47choice in 2027. Assume 14:51that for enterprise and for businesses, 14:53if that's who you're selling to, they 14:55are going to have better local compute 14:58options for their employees where they 15:00can buy laptops by the dozen. And they 15:03will be early adopters if that gives 15:05them privacy they're looking for and 15:06they will be aggressive in looking for 15:09privacy favored solutions. And that's 15:12going to affect the kinds of solutions 15:13that are built because most of the 15:14vendors out there are presuming that you 15:17will want to go through them and 15:20essentially rent cloud compute from one 15:22of the major model makers. That world 15:24may not last. Let's go to driver number 15:26five. There's an increasing market 15:29segmentation between commodity and 15:32premium AI and it's only getting 15:34reinforced. That pulse update I told you 15:36about only available for pro. If you're 15:39on plus, it's not clear when it's going 15:41to be available. I mean, presumably if 15:43they're going to pay for it with ads, 15:44they're going to run it down market 15:45eventually, but the larger trend is 15:47clear. It's clear for Claude. It's clear 15:49for Google. It's clear for Perplexity. 15:51It's clear for Open AI. Pay for what you 15:54get. The more you pay, the more you get. 15:57And so customers, whether they're 15:59individuals or employees or businesses 16:01who are willing to pay into the hundreds 16:04of dollars a month per person, are going 16:07to get access to premium human 16:10augmentation. They are going to get 16:12access to the latest models that can 16:15complete tasks for humans with a high 16:18degree of accuracy and long-term agency. 16:20and the disproportionate payoff for 16:23businesses and for people who can afford 16:26that. It's nonlinear. It's a huge deal. 16:28If you can afford premium AI and the 16:31premium AI by next year is doing tasks 16:33of four or six hours for you, tasks that 16:36take half a workday, you are going to 16:38have a tremendous advantage over 16:40everyone else in your business if 16:42they're not paying for it. You are going 16:44to be able to be in two places at once 16:47for the first time in the working world. 16:49That's massive. That is why if you are 16:52planning, you need to be planning for 16:54people to essentially duplicate 16:56themselves and be able to do multiple 16:58strands of work simultaneously. So 17:00there's roles based access control and 17:01agency questions there. There's UI 17:04questions. How do you give readr access 17:05to the agent? There are questions around 17:08purchase authority, wallet, how you 17:10check in and provide a trustful 17:11experience. At the same time, if you're 17:13going down market into the commodity 17:15space, you still have chat GPT building 17:18for a user base that is by and large on 17:22the free plan. And so even though 2% 1% 17:27maybe 5% at most of the market is super 17:30premium and pays hundreds of dollars a 17:33month, the vast majority, well over 95% 17:36is paying 20 bucks a month or nothing. 17:39And in that world, what do you do as a 17:42business to figure out your positioning? 17:44Do you want your employees to be 17:46augmented at the 20 buck a month level 17:48and you view AI as a tool add-on, or do 17:51you want them to be able to effectively 17:54duplicate themselves and do two or three 17:56times the work, but it's going to cost 17:58you more? The answer is not quite as cut 18:01and dried as it seems. If you're doing 18:03IT department planning, you have to ask 18:06yourself, is everyone at my company, if 18:08I get the budget, is everyone at my 18:10company a champion and able to adopt and 18:13use the superpowered AI the way it 18:15should be? If the new AI that we're 18:17getting in 2026, the Chat GPT6, Chat 18:20GPT7, it's like a Ferrari. Can everyone 18:22at the company drive the Ferrari? Or 18:24really, even if you gave them the keys 18:26to the Ferrari, are they really going to 18:28be able to use it and take it on the 18:29corners the way it should be taken? 18:31probably not. And this is why I keep 18:33emphasizing the the importance of this 18:3612 or 18month window in individual 18:39levelups of AI capability. This is a 18:43catch up, take the elevator to the 18:45penthouse moment for individuals and 18:48companies. It won't last forever. And 18:50the longer we have market segmentation, 18:52the more the bowling lanes in that 18:54market are going to harden. it's going 18:56to be harder and harder to jump lanes 18:58both as a person and as a company. And 19:01so when you are planning, you have to be 19:03thinking who am I marketing to? What 19:05kind of tooling do I need internally? 19:08And what is the tool set that is going 19:09to enable me to do what I need to do and 19:11what is the willingness to pay of my 19:14customer and how does that how does that 19:16imply what I need to deliver as far as 19:19premium or commoditized AI? Because you 19:21may be in a situation where your market 19:23is kind of that 95%. And as much as you 19:26want a roadmap for the cool stuff, maybe 19:28only the internal champions at your 19:30company get the fancy AI and you know 19:32your market's not going to pay for it. 19:34And so you're giving them nano banana, 19:36right? Nano banana is for everybody. 19:39That is why Google Gemini is pushing it 19:40so hard. It is an adoption play to drive 19:43familiarity with AI. That's why whenever 19:45I open Tik Tok now, I see a nano banana 19:48ad and it's always personal. It's always 19:51about change your hair color, change 19:52your style, change your vibe, take a 19:55selfie, make it look different, and now 19:57you can come up with a new look for the 19:59winter. That is the kind of thing that 20:02you have to commit to serving its scale 20:06if you want to get into the commodity 20:07space. It has to be fun. It has to be 20:09personal. It has to be quality enough 20:12that it delivers those delightful 20:14moments that are beginning to mark AI 20:16for the average consumer. But it doesn't 20:19need to have the ability to do thinking 20:22for 20 minutes and come back with an 20:23amazing PowerPoint deck. That's the 20:25premium side of things. Know your lane 20:27and know where to build. Okay, we've 20:29gone through the five key drivers. We've 20:31gone through the market segmentation 20:33piece. We've gone through technical 20:35maturation and sort of how the 20:37availability of chipsets is going to 20:39change things in 2026. We've gone 20:41through competitive velocity from AI 20:43native entrance and what that means for 20:44everybody. We've gone through how we are 20:47dealing with ROI demands in uncertain 20:50economic times and we've gone through 20:52regulatory enforcement and the 20:54regulatory market. Those are the five 20:56big drivers. I want to close by talking 20:58about a number of critical crosscutting 21:01patterns that aren't drivers themselves 21:03but that you should be thinking about as 21:06you are road mapping, building, 21:08planning, budgeting for 2026. Number 21:11one, every driver above is going to 21:13manifest differently by vertical. I 21:16mentioned the specific vertical 21:19implications in the first driver around 21:21regulatory for healthcare, for legal, 21:23for finance. That extends, right? If 21:25you're in manufacturing compliance, it 21:27looks different. If you're in robotics, 21:29it looks different. If you are in B2B 21:31SAS, it looks different. If you are in 21:33the consumer space, it looks different. 21:36you need to do the work to extend the 21:40current thinking around drivers to your 21:42space. And I'm going to prepare a prompt 21:44for you that actually helps you do that, 21:47helps you start to unpack that a little 21:49bit, and I'll include it in this post so 21:50that you can have some help as you start 21:52to think through it. I'm a big fan of 21:54prompts enabling us to kind of take 21:56these articles in new directions. The 21:58second thing I want to call out is that 22:00memory is going to look different. This 22:02is a crosscutting theme that gets at the 22:04chipset driver. It gets at the 22:06regulatory driver. What is allowed to be 22:09remembered? Who do we record it for? 22:11What are the rules around miners? Where 22:13do we have chipsets that give us more 22:15memory? One of the things that people 22:16don't realize is that memory is scaling 22:19more slowly than inference. And so 22:22inference compute is really 22:24debottlenecked right now comparatively, 22:26but memory is not. Memory is not growing 22:28as fast. We need better memory 22:33solutions. Businesses that can deliver 22:35seamless memory solutions in their area 22:38are going to have an advantage because 22:40they will close a habit loop much more 22:42effectively. That's one of the things I 22:44called out in my AI native velocity 22:47driver that incumbents have potentially 22:50an advantage on. If you can make the 22:52data you have memory, suddenly you have 22:55some stickiness there in the AI era. But 22:57memory looks different. Memory for 22:59clinics looks different than memory for 23:01legal. You have different requirements. 23:03Memory for maintenance looks different 23:05from memory for a car rental shop. 23:09Everybody has different memory needs and 23:11the startups that are building in this 23:12space are by and large not recognizing 23:14that yet. And so there's a lot of 23:16opportunity to build around the memory 23:17space whether you're an incumbent or a 23:19startup. This gets at another insight I 23:22want to call out. We've been talking a 23:24lot about verticals in this crosscutting 23:26theme. Think of that in terms of build 23:28generic horizontal tools which many 23:31vendors are pedalling right now are 23:33increasingly going to be at risk in a 23:35world when vertical expertise is 23:37becoming more defensible. One of the big 23:40crosscutting themes here is that all of 23:42these drivers reinforce the value of 23:45vertical expertise. You need vertical 23:47expertise to navigate regulatory. You 23:49need vertical expertise to navigate what 23:51customers need around local chips on 23:55their machines for their local problems. 23:57Expertise is going to matter more. 23:59Expertise is defensible and expertise is 24:02vertical tide. Which means when you are 24:04thinking about what you build this 24:06coming year, think about your vertical. 24:09Take it seriously. This may not be the 24:11year to try and jump two or three 24:12verticals at once. I know people get 24:14bold about that. 24:16This may be the year to own the vertical 24:18you're in. I will call out one more 24:19piece here. There are talent 24:21implications for the kind of world we're 24:24talking about. And if you're building 24:25for the team, a lot of people are asking 24:27me where are the engineers? Where are 24:29the architects? What does that look 24:31like? I also want to ask you about the 24:34next level of team. As you're thinking 24:35about your budgeting, don't just think 24:37about architects. Don't just think about 24:38engineers. Think about implementation. 24:41What does it take to invest in training? 24:43What does it take to invest in 24:45implementation specialists? What does it 24:47take to invest in your culture so that 24:49you have the right AI champions on the 24:51team? Not necessarily just engineers and 24:54architects who can drive the business 24:55forward. This is one of the biggest 24:58misses most businesses had in 2025. It's 25:01one of the reasons why that infamous MIT 25:05study showed up with a 95% fail rate. 25:08There's talent issues at every single 25:10business I consult at. Take the talent 25:12more seriously next year. That is also a 25:14crosscutting theme. If you get the 25:16talent right, you can negotiate these 25:18drivers more fluently because you have 25:20talent that's running faster. You have 25:22talent that knows how to replicate 25:23itself with AI. So even if you're only 25:26getting 10, 15, 20% of your team 25:29superpowered on AI and you think that's 25:32not that much, I've got news for you. 25:33That's a lot more than what most people 25:35have. Most people are struggling to get 25:37one or two% of their team superpowered 25:39on AI right now. So if you can get to 10 25:41or 20, you're way ahead. Now AI native 25:44businesses have the luxury of having 25:46everybody be AI native and superpowered 25:48but right now they're tiny and that's 25:50the risk right that's always the risk 25:52with challengers and that's always the 25:54advantage that incumbents have is their 25:56size and scale and access to capital the 25:58last thing I want to call out is that 26:00platforms are one of the biggest 26:03question marks as you think through 26:04these drivers strategically most of the 26:07things I am suggesting here argue 26:10against a strong year for platform 26:12forms. The only thing running in favor 26:16of a platform play from a road mapping 26:19build perspective is the persistence of 26:22brand. Chat GPT is a brand. Claude is a 26:25brand. Gemini is a brand. These major 26:28model makers are persisting as brands 26:31around core work primitives. If you are 26:33worried, and this is always something 26:34that comes up, so let me close with 26:36this. If you are worried about where you 26:38are going to build next year because you 26:40don't want to get run over by a new 26:42release by Sam Alman, think about the 26:45incentive sets that each of these major 26:47model makers have. Chad GPT, it's a 26:50consumer world. They're clearly moving 26:52into ads. They are building for your 26:54attention. They are building for habit 26:56loops. They want to keep you in chat GPT 27:00similar to the way that Mark Zuckerberg 27:02wants to keep you in the Meta ecosystem. 27:04They also have a strong R&D arm. The 27:06enterprise offerings they have will 27:08bring more of that R&D capability, but 27:10fundamentally they're in the business of 27:12providing very high-end compute to 27:14companies and very engaging social 27:16habits to consumers. Enthropic is for 27:19work. Enthropic is interested in 27:21providing high-end tools to people who 27:23choose to engage with them. That's the 27:24positioning for claude code. That's why 27:26they've invested so heavily in an AI 27:28model that can build Excel and 27:29PowerPoint well. And they are working 27:32specifically on connectors through MCP 27:34and work primitives. So docs, feats, 27:37making sure that the basics of work are 27:39things that you think about with clot. 27:41If you're Microsoft, you're worried. If 27:43you're building tools that are beyond 27:44primitives, it's somewhat less 27:46concerning. Finally, Google. Google 27:48needs to maintain search dominance next 27:51year in a world where ad spend is going 27:54to open up for chat GPT. Those ad 27:56budgets, are they going to come from 27:58nowhere? Everyone's going to be looking 27:59at their search revenue and asking, do 28:02we spend this much on search or do we 28:04switch this to Chat GPT? What do we do? 28:07That's going to be the question for 28:08marketers next year. And Chad GPT is 28:11going to be aggressively positioning 28:12chat as the super premium highquality 28:15intent option. Google needs an answer. 28:18That is where Gemini is going. Gemini 28:20will get into the ad space somewhere and 28:23they will start to monetize because they 28:25need it to defend Google's core search 28:28revenue dominance. There you go. That's 28:30your sneak peek at where the big players 28:32are going. I hope you don't get squished 28:34by them. Best of luck with your 2026 28:36planning. I hope this overview of the 28:38year ahead has helped you to get some 28:41clarity. That's what we're after here is 28:43clarity. And of course, use those 28:44prompts to dig deeper into your own 28:46situation. Tears.