¶ Introduction
Hi everyone. I'm Mandy and this is the AI Breakdown.
Welcome to your weekly news edition where I'll cover what happened in AI last week, why it matters all in about 15 minutes. 4 00:00:13,648.614537092 --> 00:00:21,748.614537092 Let's start with open ai because they have just launched something that's very clearly aimed at becoming the default way businesses deploy AI agents. 5 00:00:22,138.614537092 --> 00:00:24,388.614537092 A new enterprise platform called Frontier. 6 00:00:25,108.614537092 --> 00:00:26,428.614537092 The basic pitch is simple. 7 00:00:26,728.614537092 --> 00:00:31,648.614537092 You build, deploy, and manage AI agents that can do real work across your organization. 8 00:00:32,23.614537092 --> 00:00:34,903.614537092 Do it in a way that plugs into your existing infrastructure. 9 00:00:35,983.614537092 --> 00:00:41,623.614537092 OpenAI is also talking about frontier, like it's giving agents the skills you'd expect from a new human hire. 10 00:00:42,43.614537092 --> 00:00:53,293.614537092 Things like shared context onboarding, continuous learning with feedback, and define permissions so they can operate across workflows rather than just answering one-off questions in a chat window. 11 00:00:54,13.614537092 --> 00:01:00,973.614537092 OpenAI says Early adopters include HP Intuit, Oracle State Farm, Thermo Fisher, and Uber. 12 00:01:01,633.614537092 --> 00:01:06,253.614537092 With dozens more testing what it calls cross department AI coworkers. 13 00:01:07,33.614537092 --> 00:01:18,13.614537092 The quote that jumped out at me came from State Farms, Joe Park, who said, partnering with Open AI helps us give thousands of State Farm agents and employees better tools to serve our customers. 14 00:01:18,313.614537092 --> 00:01:27,108.614537092 By pairing open AI's frontier platform and deployment expertise with our people, we're accelerating our AI capabilities and finding new ways to help millions. 15 00:01:28,288.614537092 --> 00:01:34,288.614537092 So this is OpenAI planting a flag and enterprise and doing it in a very operating system for business way. 16 00:01:34,738.614537092 --> 00:01:40,468.61453709 And it's a direct shot at Anthropic, which has been deriving a big chunk of revenue from business clients. 17 00:01:40,858.61453709 --> 00:01:49,18.61453709 The fight isn't just about the best model anymore, it's about who owns deployment, governance, and the messy reality of rolling agents out safely at scale. 18 00:01:50,372.25090073 --> 00:01:54,932.25090073 And Thropic meanwhile has come out swinging with Claude Opus 4.6. 19 00:01:55,292.25090073 --> 00:01:58,802.25090073 Yes, the headline feature is its large context window. 20 00:01:59,192.25090073 --> 00:02:03,572.25090073 It's got 1 million tokens in beta, which is a big step up. 21 00:02:04,352.25090073 --> 00:02:09,122.25090073 But the more interesting bit for me personally is how explicitly anthropic is EMN. 22 00:02:09,122.25090073 --> 00:02:12,602.25090073 Beyond coding, they see Opus 4.6 23 00:02:12,602.25090073 --> 00:02:23,522.25090073 is optimized for knowledge work, so doing things like generating Excel spreadsheets, PowerPoint presentations, financial analysis, and long form research synthesis alongside code. 24 00:02:23,912.25090073 --> 00:02:34,52.25090073 They're also claiming in faster output and better reliability on long horizon tasks with a bold line that it can finish multi-day coding or analysis projects in hours. 25 00:02:34,562.25090073 --> 00:02:40,532.25090073 And then there's the new agent teams feature, which is currently a research preview inside Claude Code. 26 00:02:41,162.25090073 --> 00:02:45,2.25090073 The idea here is multiple AI agents collaborating. 27 00:02:45,122.25090073 --> 00:02:47,192.25090073 For example, one that extracts data. 28 00:02:47,582.25090073 --> 00:02:55,832.25090073 Another that writes the report coordinating automatically, so basically a divide and conquer model that looks a lot like how an actual team works. 29 00:02:56,582.25090073 --> 00:03:06,182.25090073 Andro also says this release went through its most extensive safety testing so far, including enhanced cybersecurity evaluations and better refusal handling for dodgy requests. 30 00:03:06,902.25090073 --> 00:03:13,772.25090073 In my view, the large context window and multi-agent coordination are both pointing at end-to-end business processes. 31 00:03:14,327.25090073 --> 00:03:25,127.25090073 Competitively Anthropic is clearly trying to go toe to toe with open AI's GPT five series, even talking up cases where it outperforms GPT 5.2 32 00:03:25,277.25090073 --> 00:03:26,867.25090073 in real world evaluations. 33 00:03:28,659.06908255 --> 00:03:33,639.06908255 Sticking with Anthropic for a second because they also rolled out something much more practical. 34 00:03:34,29.06908255 --> 00:03:36,159.06908255 Custom plugins for Claude Cowork. 35 00:03:36,519.06908255 --> 00:03:39,639.06908255 Cowork is their new-ish business user assistant platform. 36 00:03:39,909.06908255 --> 00:03:44,679.06908255 It's a more user-friendly sibling to Claude Cord, and now enterprises can add. 37 00:03:45,84.06908255 --> 00:03:50,934.06908255 Plugin extensions that basically program code with company tools, data sources, and rules. 38 00:03:51,384.06908255 --> 00:04:00,744.06908255 These plugins can handle departmental jobs like drafting, marketing copy, reviewing legal documents, dealing with customer support queries, or analyzing sales data. 39 00:04:01,74.06908255 --> 00:04:12,529.06908255 And the clever bit is that admins can specify which internal databases or apps the I can pull from to find steps for critical processes and even set up slash command triggers that employees can use. 40 00:04:13,779.06908255 --> 00:04:23,79.06908255 Andro also open sourced 11 of its own in-house plugins as templates, and it's making a big deal of the fact you don't need much coding knowledge to modify or build new ones. 41 00:04:23,799.06908255 --> 00:04:37,809.06908255 Their product manager, Matt Pic, summed up the goal saying, really what we're doing with this launch is just bringing plugins to cowork and giving them that kind of user friendly, UI centric flavor that will allow the maximum number of people to use them. 42 00:04:38,364.06908255 --> 00:04:44,214.06908255 For me, this is important and represents the unglamorous high ROI side of enterprise ai. 43 00:04:44,604.06908255 --> 00:04:52,314.06908255 Not just having a powerful model, but tailoring it to real business workflows so that your teams get something genuinely useful. 44 00:04:52,704.06908255 --> 00:04:54,834.06908255 But remember that governance is going to be key here. 45 00:04:55,224.06908255 --> 00:05:01,399.06908255 The moment you give agents access to live systems through plugins, you need to be on top of permission controls and auditing. 46 00:05:02,184.06908255 --> 00:05:08,244.06908255 Overall, it's a very sane step towards making AI useful to non-engineers at scale. 47 00:05:09,681.34180982 --> 00:05:18,291.34180982 Now, let's zoom out from models and tools to the money because the spending numbers last week were properly eye watering and investors did not love it. 48 00:05:18,921.34180982 --> 00:05:29,391.34180982 Reuters pulled together the scale of what's happening, and it's basically an AI spend and splurge of more than $600 billion by the biggest tech firms in 2026. 49 00:05:29,811.34180982 --> 00:05:34,881.34180982 To put that into perspective, that's nearly two billions dollars spent every day on ai. 50 00:05:35,511.34180982 --> 00:05:48,291.34180982 The two big headlines are Amazon seeing a plan's $200 billion in CapEx in 2026 and Alphabet signaling its CapEx could double this year, both heavily driven by AI data centers and chips. 51 00:05:48,831.34180982 --> 00:05:57,321.34180982 The market reaction was spiky and analysts have dubbed it the SaaS apocalypse due to a massive rotation out of traditional software into AI infrastructure. 52 00:05:57,936.34180982 --> 00:06:04,716.34180982 Alphabet stock fell around 8% at first before recovering, and Amazon dropped around 7% on the CapEx news. 53 00:06:05,76.34180982 --> 00:06:07,86.34180982 Meta dip two though less dramatically. 54 00:06:07,266.34180982 --> 00:06:17,676.34180982 And in one of those moments that tells you exactly how Wall Street thinks Nvidia surge 7% on the same news, because if everyone's buying shovels, the shovel seller tends to do. 55 00:06:17,676.34180982 --> 00:06:18,216.34180982 All right. 56 00:06:18,906.34180982 --> 00:06:21,726.34180982 What I found even more interesting is how the fear spreads. 57 00:06:22,131.34180982 --> 00:06:28,191.34180982 There's reports that investors also rotated out of traditional software and data companies on disruption fears. 58 00:06:28,731.34180982 --> 00:06:39,201.34180982 It mentions Thomson Reuters seeing record one day drops, London's, RELX, falling about 17% over a week, and Indian software exporter stocks losing $22.5 59 00:06:39,201.34180982 --> 00:06:41,91.34180982 billion in value in a week. 60 00:06:41,721.34180982 --> 00:06:46,376.34180982 Andrew Wells, CIO, at Sand Jack Investments basically summed up the vibe saying. 61 00:06:47,151.34180982 --> 00:06:55,911.34180982 The market's viewpoint is that the AI build out trade got too pricey, pulling forward all these potential future revenues and not really pricing in the risk. 62 00:06:56,331.34180982 --> 00:07:00,651.34180982 It's not that the trade is over, but it got too pricey, so it's a de-risking trade. 63 00:07:01,431.34180982 --> 00:07:08,871.34180982 So whilst this kind of CapEx wave will accelerate what's technically possible, more data centers, more chips, more AI everywhere. 64 00:07:09,171.34180982 --> 00:07:11,451.34180982 It also changes the competitive landscape. 65 00:07:12,216.34180982 --> 00:07:21,936.34180982 If the giants are building an intelligence layer into everything, smaller software companies need to be crystal clear about what they do, that a general agent can't just automate away. 66 00:07:22,566.34180982 --> 00:07:32,846.6275372 And in the short term investor expectations are getting less forgiving on the monetization front, Google made a move that I think we're going to see repeated across the industry. 67 00:07:32,846.6275372 --> 00:07:37,16.6275372 It's starting to charge for higher tier AI usage inside workspace. 68 00:07:37,601.6275372 --> 00:07:43,721.6275372 They're introducing a paid add-on called AI expanded access from March 1st, 2026. 69 00:07:44,561.6275372 --> 00:07:50,561.6275372 The way Google frames it is that basic AI features stay included in most business and enterprise plans. 70 00:07:50,831.6275372 --> 00:07:56,171.6275372 So everyday stuff like help me write in Gmail or summaries and docs doesn't suddenly vanish. 71 00:07:56,291.6275372 --> 00:08:02,291.6275372 But if you are a power user who wants higher usage limits on the more advanced features, you'll need the add-on. 72 00:08:03,191.6275372 --> 00:08:05,351.6275372 Whilst I'm not convinced that means the free trial. 73 00:08:05,351.6275372 --> 00:08:06,701.6275372 AI era is ending. 74 00:08:07,151.6275372 --> 00:08:12,821.6275372 Google's latest move signals a transition to a more disciplined usage-based model for the enterprise. 75 00:08:13,451.6275372 --> 00:08:21,446.6275372 Moving over to hardware now, Cerebra has just raised a frankly enormous $1 billion series H round, putting it at about a $23.1 76 00:08:21,446.6275372 --> 00:08:26,286.6275372 billion post-money valuation, nearly tripling its valuation in a bit over four months. 77 00:08:28,185.51236838 --> 00:08:34,455.51236838 Cerebra is one of the more interesting NVIDIA challenges because it's not just making a slightly different GPU. 78 00:08:34,815.51236838 --> 00:08:43,5.51236838 Its thing is wafer scale engines, basically the world's largest single silicon chips using an entire wafer as one giant processor. 79 00:08:43,665.51236838 --> 00:08:54,870.51236838 The research notes these WSE chips are 56 times bigger than a typical GPU, and the latest WSE three has more than 850,000 core on one wafer. 80 00:08:55,785.51236838 --> 00:09:02,475.51236838 The Round was led by Tiger Global with backers including Benchmark Chu and strategically A MD as well. 81 00:09:03,45.51236838 --> 00:09:04,935.51236838 And there's one detail that really matters. 82 00:09:05,55.51236838 --> 00:09:10,395.51236838 Just weeks before the fundraise open did a deal to procure Cerebra hardware for model inference. 83 00:09:10,725.51236838 --> 00:09:15,855.51236838 If Open is willing to run inference on your kit, that's a serious validation for me. 84 00:09:15,885.51236838 --> 00:09:19,515.51236838 This means we're watching the picks and shovels phase of the AI boom. 85 00:09:19,845.51236838 --> 00:09:20,985.51236838 Play out in real time. 86 00:09:21,420.51236838 --> 00:09:27,990.51236838 Nvidia is still the default, but the demand for compute is so high that credible alternatives get a lot of attention. 87 00:09:28,20.51236838 --> 00:09:35,460.51236838 And funding more competition here could mean better availability and eventually pricing, which helps everyone downstream. 88 00:09:36,90.51236838 --> 00:09:38,220.51236838 The pressure now of course, is execution. 89 00:09:38,580.51236838 --> 00:09:40,500.51236838 It's one thing to build mind blowing silicon. 90 00:09:40,920.51236838 --> 00:09:46,200.51236838 It's another to build the ecosystem and the sales machine that convinces enterprises to switch. 91 00:09:47,757.78509565 --> 00:09:58,527.78509565 Finally, this year's Super Bowl felt like a bit of a turning point for AI advertising, not because AI appeared for the first time, but because the messaging got sharper and more openly contested. 92 00:09:59,247.78509565 --> 00:10:03,627.78509565 One thing that stood out, the AI companies weren't just showing off features. 93 00:10:03,777.78509565 --> 00:10:12,627.78509565 They were competing on positioning and thropic run a cloud ad that basically made one clean argument that ads don't belong inside AI conversations. 94 00:10:13,527.78509565 --> 00:10:21,852.78509565 It's hard not to read that as a direct swipe at the whole idea of ad supported chat and by extension, an obvious swipe at OpenAI. 95 00:10:23,7.78509565 --> 00:10:30,777.78509565 OpenAI, on the other hand, leaned into the build with the eye story that you can just make stuff now and put the emphasis on tools like Codex. 96 00:10:31,842.78509565 --> 00:10:34,692.78509565 Second, I wasn't limited to the AI ads. 97 00:10:34,722.78509565 --> 00:10:35,592.78509565 It was everywhere. 98 00:10:36,162.78509565 --> 00:10:46,152.78509565 TechCrunch pointed out that this year pushed last year's trend further with brands using AI, both as the thing they're selling and as part of the look and feel of the commercial itself. 99 00:10:46,722.78509565 --> 00:10:53,322.78509565 And Business Insider went further, seeing nearly a quarter of the ads either referenced AI or used the AI tools. 100 00:10:53,802.78509565 --> 00:10:56,232.78509565 So AI is starting to become a default. 101 00:10:56,232.78509565 --> 00:11:00,372.78509565 Language brand's reach for whether they're selling software phones. 102 00:11:00,762.78509565 --> 00:11:06,762.78509565 Oh yes, vodka third, the backlash is real, and viewers are basically playing spot. 103 00:11:06,762.78509565 --> 00:11:17,82.78509565 The ai, the Verge Roundup was pretty blunt about it, highlighting a notable jump in AI generated or AI looking ads, and the reception wasn't great. 104 00:11:17,472.78509565 --> 00:11:22,572.78509565 A lot of it came off as cheap, sloppy, and creatively thin, and there's a weird meta effect now. 105 00:11:22,872.78509565 --> 00:11:29,262.78509565 Even ads that weren't AI made got accused of being AI just because the visuals felt a bit too synthetic. 106 00:11:30,57.78509565 --> 00:11:30,897.78509565 You can't win. 107 00:11:31,497.78509565 --> 00:11:33,627.78509565 That feels like a genuine cultural shift. 108 00:11:34,17.78509565 --> 00:11:38,607.78509565 Gen AI aesthetic is starting to register as a warning sign, not a flex. 109 00:11:38,967.78509565 --> 00:11:42,357.78509565 And finally, the measurement game is changing in real time. 110 00:11:43,107.78509565 --> 00:11:53,397.78509565 Axios reported that X rolled out a real time Super Bowl ad tracker called Brand Ranks measuring Engagement and Sentiment as the ads aired, which tells you what the new Super Bowl meta is. 111 00:11:53,952.78509565 --> 00:12:06,72.78509565 It's not just win the room, it's win the timeline and ideally win it with the right emotion, not just attention, but for me, the Super Bowl has a funny way of doubling as a market sentiment indicator. 112 00:12:06,852.78509565 --> 00:12:14,172.78509565 When a new tech wave is so confident it's buying the most expensive, most public ad inventory on earth, that's often a sign. 113 00:12:14,172.78509565 --> 00:12:15,432.78509565 We're late in the hype cycle. 114 00:12:15,732.78509565 --> 00:12:18,882.78509565 The we've made it moment right before reality shows up with a bang. 115 00:12:19,542.78509565 --> 00:12:24,522.78509565 The classic example is this, so-called dot com, super Bowl in January, 2000. 116 00:12:24,882.78509565 --> 00:12:27,552.78509565 When a huge share of the ad slate was.com 117 00:12:27,552.78509565 --> 00:12:36,492.78509565 brands spending millions to look inevitable and the bubble deflated soon after, with several of those advertisers folding within the year, and we've seen a modern rerun. 118 00:12:36,942.78509565 --> 00:12:45,42.78509565 Super Bowl 2022 got nickname the Crypto Bowl because exchanges and platforms splashed out on big, glossy ads. 119 00:12:45,342.78509565 --> 00:12:50,592.78509565 And then the sector slid into crypto winter with major players like FTX collapsing later that year. 120 00:12:51,432.78509565 --> 00:12:57,672.78509565 The point isn't that Super Bowl ads cause crashes, it's that they're allowed expensive symptom of peak confidence. 121 00:12:57,972.78509565 --> 00:13:08,442.78509565 When an industry starts paying eight figures for 30 seconds just to announce we are, the future history says you should at least check whether you're watching innovation or a warning signal. 122 00:13:09,252.78509565 --> 00:13:14,772.78509565 Maybe we are in an AI bubble in pockets, but the optimistic case is that this isn't just hype. 123 00:13:15,147.78509565 --> 00:13:28,347.78509565 Unlike past manias, a lot of the spend is building real reusable infrastructure chips, data centers, software platforms, and it's already translating into meaningful revenue for key players, which suggests real substance. 124 00:13:28,347.78509565 --> 00:13:36,237.78509565 Under the story, time will tell of course, but I'm in the mindset of AI is here for the long run, regardless of short-term volatility. 125 00:13:37,831.42145929 --> 00:13:39,991.42145929 That's all for this week's AI roundup. 126 00:13:40,381.42145929 --> 00:13:44,101.42145929 If you found value in this breakdown, please leave a rating and hit subscribe. 127 00:13:44,371.42145929 --> 00:13:45,301.42145929 See you next week.