AI's new power plays: from personal agents to power grids.

Meta’s launch of its "Muse" personal AI assistant stands out as a statement of intent in Big Tech’s battle to dominate everyday AI. Rolling out to a broad U.S. user base, **Muse** can autonomously handle tasks like managing emails, calendars, travel bookings, shopping, and more across various connected apps and services ([1]). This is not just another chatbot – it’s an AI agent that lives across Meta’s ecosystem (accessible via a dedicated app, web, or WhatsApp) and even on upcoming AR glasses, with a freemium model offering basic service and $20 or $100 monthly tiers for power users ([2]). By leveraging the **$130 billion** Meta plans to invest in AI infrastructure this year ([3]), CEO Mark Zuckerberg is betting on a "personal superintelligence" strategy to embed AI deeper into users’ daily lives and finally move beyond the confines of social feeds.

However, this new convenience comes with major strategic questions about trust and control. To truly assist users, **Muse** requires unprecedented access to personal data and accounts – effectively asking consumers to hand Meta “the keys to their inboxes, payments, and calendars” ([4]). That might be a tough sell given Meta’s checkered privacy history: *Reuters* reports even some insiders worry the agent could mis-handle sensitive data, and the debut comes on the heels of an $18 billion settlement over user harms on its social platforms ([5]). The flipside is that if Meta can overcome these trust issues, it could position Muse as an indispensable digital butler woven through its massive user base on Facebook, Instagram, and WhatsApp. That "agentic AI" vision – where consumers delegate real world tasks to AI – puts Meta in direct competition with **OpenAI**, **Google**, **Apple**, and others vying to be the go-to AI assistant across devices and services ([6]).

Notably, each tech giant is approaching the personal AI race from its own strategic angle. **Apple**, for instance, used its latest iPhone and Watch launch this week to showcase a revamped *Siri AI* that, while less overtly ambitious than Muse, integrates generative AI across nearly all new devices ([7]). Apple’s version leans on privacy and a unique two-tier architecture: on-device models for quick tasks and a “private” Apple cloud for heavier reasoning ([8]), with usage limits and plans to charge for premium AI features down the line ([9]). That strategy signals Apple’s intent to offer smarter assistants without sacrificing its brand advantage in privacy – a potential differentiator as users grow wary of giving companies full access to their digital lives. Meanwhile, Google has so far played its consumer AI cards close to the vest, but industry chatter hints at forthcoming *Gemini* AI services and the company’s existing prowess in Android and Assistant gives it significant distribution. The overarching trend is clear: owning the **platform for personal AI agents** – and the rich user data and engagement that come with it – may define the next era of consumer technology competition.

The past two days also saw a flurry of moves to embed AI deeper into specific industries, as both startups and incumbents seek defensible advantages via proprietary data and domain expertise. A prime example is **OpenAI’s new ChatGPT for Financial Services**, announced on September 10. This is a tailor-made version of ChatGPT designed for bankers and analysts, combining OpenAI’s latest model *GPT-6 “Astra”* with live, premium financial information. Through partnerships with firms like Morgan Stanley and Evercore during development ([1]), OpenAI integrated a suite of market data from providers such as LSEG (London Stock Exchange Group) News, PitchBook, Daloopa, Quartr and Crunchbase, which comes pre-loaded into the platform ([2]) ([3]). In effect, OpenAI is bundling raw intelligence (its most powerful model) with valuable financial content and workflow integrations, offering an all-in-one research and modeling assistant for the finance sector. This shift transforms ChatGPT from a general-purpose assistant into more of an **enterprise software product**, complete with data access, compliance guardrails and template outputs for financial analyses.

Strategically, OpenAI’s vertical push into finance blurs industry boundaries and puts the AI leader on a collision course with both incumbent providers and AI startups. In the $30+ billion financial data and analytics market long dominated by terminals and platforms like **Bloomberg**, FactSet, and Reuters, OpenAI is now positioning itself as a new kind of competitor – an AI-first research suite that could augment or displace traditional tools. By hosting third-party data within its own system (and providing features like source citation tracking for due diligence) ([4]), OpenAI is moving “up the stack” to capture more of the value around its models. The stakes are high: *Reuters* notes that *GPT-6 Astra* is not just more powerful than previous models but has been re-tooled for business uses, and finance is an early beachhead where Microsoft, Google, and a wave of finance-specific AI startups are all racing to plant their flag ([5]). This step by OpenAI will force specialized fintech AI vendors to rethink their roadmaps – and it’s a clear signal to every industry that today’s AI platform providers may not stay just in the background infrastructure role. If critical data or domain workflow integrations can create a competitive moat, expect AI players to move there quickly, potentially competing with their own ecosystem partners.

Incumbent enterprises are likewise making bold plays to secure AI advantages in their sectors, often via acquisitions and alliances. This week brought news that **Salesforce** is in talks to acquire *Listen Labs*, a fast-growing AI-driven customer research platform, for around $2 billion ([6]). Listen Labs boasts a global panel of 50 million consumers and an AI-powered system that conducts qualitative interviews and analyzes feedback at scale ([7]). If it proceeds, the deal would follow Salesforce’s recent $3.6B purchase of **Slack** (for AI-enhanced customer service automation) and highlight a pattern: big companies buying AI-native firms to gain unique data assets and capabilities that can’t be easily replicated. In the media realm, AI startup **Suno** announced licensed music-generation models built with record labels like Warner Music and BMG, retiring its unlicensed predecessors ([8]). In exchange for providing training data and legal cover, those labels will receive a share of subscription revenues from Suno’s 2 million-plus paying users ([9]). This kind of partnership-driven model innovation could allow AI challengers to outmaneuver incumbents by solving thorny issues (like intellectual property rights) while rapidly scaling a user base. The common thread: whether through acquisition or partnership, companies are moving to lock in **proprietary data pipelines and distribution channels** that turn AI capabilities into lasting strategic advantages.

Another lesson from this week’s developments is that the AI race is increasingly waged on the invisible battleground of hardware and energy. The clearest example came from **Amazon** and **Qualcomm**, which revealed a multi-year partnership to co-develop custom AI chips for Amazon’s cloud (AWS) data centers ([1]). Rather than relying solely on off-the-shelf GPU accelerators from Nvidia, Amazon Web Services is collaborating with Qualcomm to design its own silicon for large-scale AI inference – the work of running AI models for millions of users in real time. This deal could span several generations of processor and networking technology ([2]), signaling that Amazon is serious about matching Google and Microsoft in building a diversified chip portfolio for AI. The strategic rationale is cost and control: as AI usage grows, serving complex models to vast user bases becomes extraordinarily expensive – in fact, operating costs for popular AI services can exceed the cost of training the models themselves ([3]). By designing tailored chips and advanced optical interconnects for its data centers, AWS aims to both improve performance and dramatically lower the per-query cost of AI, undercutting a key advantage of Nvidia and reducing dependency on a single supplier ([4]).

Competition for core AI infrastructure is extending into unprecedented territories – even energy. On September 9, **Google** announced a commitment of roughly **€13 billion** to expand its AI computing footprint in Finland, including new data centers at four sites and a first-of-its-kind 22-year power purchase agreement for nuclear energy ([5]). Under the deal with utility Fortum, Google will eventually buy up to 50% of the output of Finland’s Loviisa nuclear plant over the 2030–2049 period ([6]). This long-term contract will effectively keep that reactor (which supplies about 10% of Finland’s electricity) running two decades longer than planned ([7]), ensuring Google exclusive access to stable, carbon-free power through mid-century. The move underscores how vital reliable electricity has become to AI leadership: the *next phase of the AI race may be constrained as much by power and grid capacity as by model architecture* ([8]). By locking down a secure energy supply, Google is not just investing in servers – it’s preemptively solving one of the biggest bottlenecks to scaling AI. This raises the bar for other cloud providers (and even large enterprises with AI ambitions) to consider similarly creative strategies for **securing critical resources**, from energy to specialized hardware and networking.

Finally, several of the most high-profile AI-native startups are fortifying their defenses in ways that should give incumbents pause. This week saw **Harvey**, a two-year-old AI company focused on legal services, close a $550 million funding round that values it at roughly $15.6 billion ([1]). Harvey has already signed up 3,000+ enterprise customers – including 80% of the top 100 law firms in the U.S. and even multiple Fortune 10 corporations – and surpassed $400 million in annual recurring revenue ([2]). Those numbers reflect a remarkable pace of adoption in a traditionally conservative industry, and investors are imputing a sky-high 39× revenue multiple in anticipation that Harvey will transform the practice of law ([3]) ([4]).

Crucially, Harvey is not simply riding on the coattails of bigger AI platforms – it’s actively building its own. The startup’s leadership argues that “all software companies need to turn into AI companies, full stop,” and is using its new war chest to develop proprietary AI models rather than remain dependent on general-purpose systems from OpenAI or Anthropic ([5]). In fact, Harvey’s first in-house model, **Harvey Tenet**, was built by adapting an open-source 2.8-trillion-parameter model from a research lab in Beijing, illustrating how top-tier talent will leverage *any* available frontier technology to gain an edge ([6]) ([7]). Harvey is also on an acquisition spree – four so far in 2026 – including a smaller AI “guardrails” startup that stress-tests autonomous agents for safety and reliability ([8]). This aggressive strategy is a bid to own the full stack of technology and trust, which could set the firm apart as legacy legal tech providers (and even its own model suppliers-turned-competitors, as OpenAI and Anthropic build legal AI features) try to catch up ([9]).

And it’s not just legal tech seeing a startup surge. In the software development arena, AI coding assistant company **Cognition AI** raised $2 billion in new funding at a stunning $48 billion valuation ([10]). Even more telling was the traction justifying that price: Cognition’s autonomous coding agent (a potential rival to Microsoft’s GitHub Copilot) has seen its revenue run-rate rocket from $492 million in May to nearly $900 million by early September ([11]). This suggests AI-driven code generation might be one of the first truly massive markets for enterprise AI, as companies invest to boost developer productivity at scale ([12]). But sustaining such momentum will require fending off competition on multiple fronts. Cognition’s $48 billion price tag assumes it can become a dominant platform for AI-assisted software development – rather than just a clever feature that bigger cloud and tooling providers replicate ([13]). Likewise, Harvey’s long-term success depends on whether controlling the model layer will protect its lead, or simply invite new challengers in a field projected to grow nearly 8× to $41 billion globally by 2034 ([14]). For executives across industries, the takeaway is that AI-native entrants are scaling at unprecedented speed, and they’re not content to be features in someone else’s ecosystem. The time for incumbents to decide whether to partner, acquire, or compete with these upstarts is fast running out.