Loading summary
Transcript1 lines
- [00:00]
A
Hi everyone, I'm Andy and this is the AI Breakdown. Welcome to your weekly news edition, where I'll cover what happened in AI last week and why it matters. Let's start with the Pentagon, because this story says quite a lot about where AI procurement is heading On 1 May, the Department of Defense awarded contracts that let eight AI companies deploy models inside systems classified at secret and top secret levels. The names on the list were OpenAI, Microsoft, Google, AWS, Nvidia, SpaceX, and Reflection AI. SpaceX is an interesting one as it doesn't actually have a major frontier model of its own, so that inclusion looks like a bet on secure infrastructure and XAI ties as much as model quality. Oracle appears on a separate but related award. Unsurprisingly, the name missing was Anthropic, given their ongoing legal battles with the Pentagon. We covered Anthropics and Junction Fight in previous episodes and also the Mythos Cyber story as a possible path back into government work. However, this week has made that route appear much narrower. The DoD contracts are open ended vehicles with multi year ceilings the Pentagon hasn't disclosed. Bloomberg Government estimates classified federal AI spend at about $8 billion to $12 billion over the next five years. Anthropic is now excluded from that, at least in this segment of the market. What's intriguing is that the Pentagon didn't select just one winner. It intentionally chose multi vendors. This reflects an important insight about the U.S. government. The largest serious buyer in this category doesn't want to rely solely on a single model provider. It prefers optionality, bargaining power, and perhaps a bit of political leverage too. Reflection AI's inclusion is a part of this strategy. It's a 2024 startup founded by ex DeepMind researchers focusing on automated software engineering and large model workflows, and its inclusion seems like a hedge against just handing everything to the obvious giants. My view is that this doesn't just hurt Anthropic and Washington. It tells every Frontier lab that in the next phase of AI alignment with government priorities can matter as much as raw model quality. That's a big shift. It means procurement is becoming a policy filter, not just a technical evaluation. Being labeled a supply chain risk is undeniably negative, and you certainly don't want that kind of association. However, there's also an angle where commercial buyers in healthcare, finance and parts of Europe hear a more straightforward message. Anthropic was the lab that stood firm on not allowing autonomous targeting and mass surveillance in certain sectors. That stance is not a weakness, but rather a positioning. So this reshapes the federal leaderboard, but also underscores that policy disagreements now have a direct impact on market access. And that's a larger story than one contract list. This episode is sponsored by Notify Technology, a health and safety software platform that helps teams handle incident reporting, audits, inspections and risk assessments in one place. And with Notify Spark, its AI companion, teams can speed up incident management and investigations while cutting admin. Learn more@notifytechnology.com the other significant story this week is OpenAI ending its old Azure only period. On 27 April, OpenAI and Microsoft announced a reworked commercial agreement. Microsoft called it the next phase of the partnership, which is corporate language doing a lot of work in practice. The key points were Microsoft's revenue share gets capped, the previous AGI clause goes away, which would have empowered OpenAI to cut off Microsoft's access to its technologies. Upon achieving Artificial General Intelligence, Microsoft obtains a non exclusive IP license for OpenAI models and products through 2032, and OpenAI is no longer effectively tied to Azure alone. As Satya Nadella stated in Microsoft's announcement, Azure remains central, but both companies are now free to collaborate with others. Then, a little less than a day later, the next phase was revealed and WS&OpenAI announced that GPT 5.4 would go live in preview on Bedrock immediately, with GPT 5.5 following shortly. Codex is also available and AWS has added Bedrock Managed agents to facilitate building agents on top of OpenAI's reasoning models for AWS first organizations. This is significant. Before this opened up, One of the main challenges with OpenAI in large enterprises was the cloud politics. If your system was on AWS or your governance and IAM stack lived there. Having to route substantial OpenAI usage to through Azure added friction. Now that barrier is gone. AWS accounts include the majority of large US enterprises, so this opens OpenAI to a customer base it had made unnecessarily difficult to access. Bedrock provides AWS customers a way to use GPT models within an environment they are already familiar with, with controls they trust. This shift changes the market in several ways. Firstly, it boosts access as we just covered. Secondly. Secondly, it alters pricing dynamics, with OpenAI pitting Azure and AWS against each other. Analysts expect pricing pressure to emerge quickly, with potential API price reductions within about 90 days. Thirdly, it confirms what many Microsoft and OpenAI are no longer acting as a locked pair. They remain close, but in the way large companies do when they're also gearing up for life without the other. This aspect is as significant as the bedrock headline, Microsoft has already been building more flexibility with its own models, which we discussed earlier this month, and now the contract structure aligns with the strategy, OpenAI gains more routes to market, Microsoft keeps valuable IP rights and more freedom, and AWS can portray itself as the neutral marketplace where OpenAI, Anthropic, Mistral and its own tools can all coexist. Now let's switch to the financials because the hyperscaler earnings this week were frankly extraordinary. On 29 April, Alphabet, Microsoft and Amazon all reported within hours of each other and the capex numbers pointed to what may be the biggest single year infrastructure buildout the tech industry has ever seen. Alphabet doubled its 2026 CapEx guidance to between 180 and $190 billion. Microsoft guided to around $190 billion for 2026 and CFO Amy Hood said most of that goes on CPUs and GPUs. Amazon had already spent $43 billion by 31 March and is projecting roughly $200 billion for the full year. Admeta's updated guidance of 1:25 to $145 billion and you're looking at about $700 billion across the big four. That's the headline number. Some analysts have it higher. The more interesting question is what it means first, it pushes back hard on the AI bubble charter from a few months back. And whilst the tension is still there between building too many data centers and not enough, this kind of funding will ease compute scarcity, not eliminate it. Every hyperscaler is effectively over provisioning, which should make 2026 and 2027amore favorable buyer's market for inference capacity than 2024 or 2025. There's also a second narrative within the first the custom silicon story. Andy Jassy noted AWS's chips business is now generating over $20 billion in revenue, growing rapidly. Analysts at Bernstein and Bank of America openly regard hyperscaler custom chips as as a potential threat to Nvidia's dominance over the 2027-2029 period. With solutions like Trainium, Google's TPUs, Meta's Mtium and Microsoft Maya, the hyperscalers are building genuine alternatives, not perfect substitutes, but viable options. This is where things get practical. Over the next couple of years, enterprises will evaluate not only model performance but but also the economics of model hosting. Cranium hosted Claude, TPU hosted Gemini, Nvidia hosted GPT. The decision regarding model and infrastructure is becoming increasingly intertwined. The Motley fool deemed this the largest technology build out in history, which may sound dramatic but is probably justified in this case. In my view, AI infrastructure is beginning to resemble essential IT expenditure similar to electricity or telecoms. Not glamorous, just foundational. The challenge is that all this spending creates a future depreciation burden. Roughly speaking, spreading $700 billion over a five year useful life result in approximately $140 billion annually in depreciation from 2027 onwards. So operating margins could face pressure unless AI revenue keeps pace for buyers, though in the near term the impact is positive. Increased supply typically means better availability, improved economics and more negotiating leverage than in the past two years an intriguing story in the enterprise sector came out of Europe this week. On 28 April, Mistral launched Workflows in public preview. It's an orchestration engine for long running AI processes and the bigger point is that Mistral thinks that the hard part in enterprise AI now isn't the model, it's making the thing run reliably inside real business processes. Workflows is built on temporal a practical choice given that this durable execution layer already supports orchestration at companies like Netflix, Stripe and Salesforce. Mistral's message is clear. The control plane operates with Mistral, but the workers and data processing remain within the customer's environment, whether on cloud, on prem or some hybrid setup. For European buyers concerned about data residency, GDPR and the AI act, this deployment strategy is a significant part of the appeal. The launch clients support this narrative ASML, CMA, CGM, Le Bancpostale, France travail, and also a banker and MoEV. These are not random names. They are the types of organizations concerned with sovereignty, auditability and avoiding the transfer of sensitive process data to an external cloud unless it's unavoidable. Mistral also claims Workflows was already handling millions of executions daily during pre launch testing, highlighting where it wants to position itself in the technology stack. My perspective is that the margin is shifting up the stack. The model is becoming the cheap part, or at least the less differentiated part. The expensive bit is getting AI to behave predictably inside a real business process. That's why orchestration matters. It's where reliability, compliance and switching costs start to live. Another story that's hard to overlook is ineffable intelligence. On 27 April, the London startup announced a $1.1 billion seed round at a $5.1 billion valuation, which is an astounding figure for a company that's only been around for a few months. It's the largest seed round ever completed in Europe. But the founder is David Silver, the former DeepMind researcher behind AlphaGo, AlphaZero and AlphaStar, so investors are evidently placing their faith in him and his research direction rather than any present product. TechCrunch summed up the central idea well Silver aims to develop a system that learns via reinforcement learning, deviating from primarily imitating human data. He's framing the mission as creating a super learner that derives knowledge from its own experiences, which marks a distinct departure from the typical scale up of today's LLMs. This matters more for the market's shape in the coming years. Sequoia, Lightspeed, Nvidia, Google DST and even the UK Sovereign AI Fund are essentially hedging against the current consensus. My take is that this isn't really a funding story, it's a rebellion story. When over a billion dollars goes into a lab with no product and a thesis that cuts against the dominant LLM playbook, that's the market saying it no longer believes OnePath will own the future. To me, that's the real signal. Right, let's quickly touch on a few more before we close. Firstly, you've probably been hearing a lot of back and forth about the EU AI act, but there was a significant breakdown in Brussels recently that hasn't quite hit the mainstream headlines yet. The omnibus talks, basically the sessions meant to clean up the final rules and potentially move some deadlines around, are collapsing. They still haven't reached a consensus, and for anyone in the business world, that is a much bigger deal than it sounds. A lot of people were banking on these talks to push back the compliance deadlines. Now that the talks are stalling, that safety net is disappearing fast. The August 2 deadline for high risk AI systems is staying exactly where it is. So what does that actually mean for you? If your company uses AI for things like scanning resumes, evaluating creditworthiness, or even certain types of employee monitoring, the EU classifies those as high risk. Because these talks failed, the clock hasn't stopped. Any high risk system you're using or launching needs to be fully compliant by August 2, or you're essentially operating outside the law. The risk here isn't just a slap on the wrist. We're talking about potential fines that can scale up to 3% of your global annual turnover. However, for prohibited practices, it can be as high as 7% or 35 million euros. But even beyond the money, there's a massive operational risk. If you aren't compliant, risk regulators can literally pull your product from the market or force you to shut down the system. If that AI is a core part of how you hire people or manage your operations, that's a day zero event for your business. Anthropic also introduced Claude Security in public beta for enterprise clients. This code based security product, powered By Claude Opus 4.7, scans for vulnerabilities, confirms findings to reduce false positives, and provides patch suggestions following the Pentagon exclusion. This feels like a strategic message. They might be excluded for now from classified defence contracts, but the enterprise software market is still accessible. Meta had a staggering statistic to share on its earnings call. The company revealed its business AI tools are now handling about 10 million conversations weekly, up from around 1 million at the year's start. This is a tenfold increase in approximately 90 days, mainly through WhatsApp and Messenger, for small to medium sized businesses. Meta claims it expanded the product in Q1 across regions such as Latin America, Indonesia and parts of Asia Pacific, so it's a free for now monetization layer to play something that should give customer service software vendors pause. Xai, meanwhile, introduced the Grok 4.3 API. Key features included a 1 million token context window, native video input and roughly a 40% input price reduction. That price move is perhaps the most significant aspect of this announcement, as it adjusts expectations within the agent tier API market and pressures others to react and one more development to monitor. ODA has undertaken its largest strategic pivot in years, expanding beyond meeting notes into enterprise search across Gmail, Drive, Notion, Jira and Salesforce, using MCP as the integration layer. TechCrunch hints that Outlook teams, SharePoint and Slack are next, revealing the broad scope of this change. That's all for this week's AI Roundup. If you found this breakdown valuable, please leave a rating and hit subscribe. See you next week. Sam.