The money got the headlines today. AMD agreed to build two gigawatts of AI computing for Anthropic and to invest up to $5 billion back into its own customer, one of a string of deals that put the industry's capital spending on open display: OpenAI lifted its infrastructure plan to $750 billion, and Alphabet raised its 2026 capex guidance high enough to knock 6% off its stock even as revenue and cloud sailed past estimates. Underneath the balance-sheet drama, the products aimed at operators kept shipping — OpenAI opened an enterprise platform for voice and chat agents, Cursor launched a model router that cuts cost 60%, and Monday.com cut a fifth of its staff to reorganize around AI. Here is what matters and why.
AMD and Anthropic sign one of the largest chip deals yet
On July 22, AMD and Anthropic announced a strategic partnership under which Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs, with the first 1-gigawatt tranche coming online in the first half of 2027. The chips will ship inside AMD's Helios rack-scale systems — MI455X accelerators paired with EPYC "Venice" CPUs, Pensando networking, and the ROCm software stack — and build on MI355X hardware Anthropic already runs. AMD separately committed to a strategic equity investment of up to $5 billion in Anthropic, structured to be made over time as deployment milestones are met rather than paid upfront.
The deal is notable on two fronts. First, it is a second frontier lab publicly diversifying away from Nvidia: Anthropic co-founder Tom Brown framed it as being able to "map the right workloads to the right hardware" and lock in scarce capacity for Claude. Second, it is circular in a way that has become the signature of this cycle — the chip vendor invests in the customer that buys its chips, a pattern that echoes Nvidia-OpenAI and now underwrites demand with the supplier's own balance sheet. AMD CEO Lisa Su said the partnership pairs "Anthropic's leadership in frontier AI with the full strength of AMD high-performance computing." The two also signed a multi-year engineering agreement to use Claude to optimize workloads and accelerate ROCm development, and AMD will adopt Claude across its engineering teams. For operators, the signal is supply: a second credible, at-scale GPU vendor for the model that powers a growing share of business tooling reduces single-vendor risk and, eventually, price.
The compute bill comes due
The AMD deal landed the same week the sector's spending numbers reset upward across the board. OpenAI reportedly raised its infrastructure plan to $750 billion through 2030 — roughly a 25% jump from its earlier estimate — and is taking more direct control of construction as its original Stargate effort stalls. Its flagship project, a $20 billion campus in Georgia dubbed Project Camellia, spans 1,400 acres and will draw at least 3.2 gigawatts, with most of that capacity — about 5.8 gigawatts across the broader footprint — coming from natural gas. OpenAI recently hired Brett Mayo, who built xAI's Colossus site in Memphis, to run the buildout.
Alphabet put hard numbers behind the same trend. In Q2 results reported July 22, revenue rose 24% to $103.62 billion and Google Cloud grew 82%, but management raised full-year capital-expenditure guidance to $195–205 billion from $180–190 billion, and the stock fell about 6% despite the beat — a sign investors are starting to scrutinize the return on all this spending. CEO Sundar Pichai said nearly 90% of the Fortune 100 now use Gemini Enterprise and that Google is "applying a lot of our compute and effort" to its next flagship, Gemini 4. Meanwhile TSMC, which fabricates most of these chips, committed an additional $100 billion to its Arizona operations, lifting its total U.S. investment pipeline to roughly $265 billion. The through-line: the AI supply chain is pouring capital in years ahead of the revenue meant to justify it, and the market is no longer applauding unconditionally.
OpenAI Presence and the enterprise agent land grab
OpenAI launched Presence on July 22, an enterprise platform for deploying and managing real-time voice and chat agents under company-defined guardrails. It bundles the plumbing that has kept most pilots from reaching production: policies and standard operating procedures, permission controls limiting what data and systems an agent can touch, escalation rules that route to humans, pre-deployment simulations that stress-test agents against edge cases, and a Codex-powered improvement loop that reviews live sessions and proposes changes staff must approve. OpenAI says Presence already runs its own English-language phone support and resolves 75% of inbound issues without a human, and it named BBVA Mexico as an early customer. It is available only through a limited program staffed by OpenAI's Forward Deployed Engineers — not self-service — with no public pricing.
Presence does not arrive in a vacuum. The enterprise agent-platform category now has four heavyweight entrants in a single month: Presence, Google Cloud's expanded Gemini Enterprise, Meta's Business Agent Platform, and the Nvidia–ServiceNow "Project Arc" partnership. Each pitches the same buyer with the same promise — deploy fleets of agents across the business with governance, permissions, and audit built in. The competition is good news for operators: it standardizes the hard parts (evaluation, escalation, monitoring) that individual teams have been reinventing, and it signals that vendors now believe the blocker to agent adoption is trust and control, not raw model capability.
Model and tooling news
Cursor shipped Cursor Router, a system that automatically routes each request to the model best suited to it. Early-access customers saw frontier-quality output at roughly 60% lower cost and a lower cost per commit than sending everything to Opus 4.8, with no quality drop-off. It is live on Teams and Enterprise plans, and admins can allow or block specific models and set defaults — a concrete example of the "put the expensive model only where it earns its keep" economics that increasingly define AI budgets.
Anthropic drew separate attention. Prediction markets spent the day betting on a Claude Opus 5 launch — the model has been spotted in Cursor under the "Honeycomb" codename with a 1M-token context window — but as of this writing Anthropic has made no official announcement, and earlier rumored release dates passed without one; treat it as unconfirmed. Anthropic did ship "Record a Skill" in Claude Cowork, which lets users teach Claude a workflow by recording and narrating their screen, then replays it as a repeatable skill (Max, Pro, and Team tiers). And Microsoft and Mistral expanded their partnership on July 21, bringing Mistral Medium 3.5 and OCR 4 into Microsoft Foundry and Medium 3.5 into Copilot Studio, backed by a multibillion-dollar expansion of Europe-based GPU capacity and deployment options that reach fully disconnected environments — aimed squarely at regulated industries. Separately, the U.S. Department of Energy and Arcee AI unveiled Genesis-Science-1, an open-weight model for scientific computing, with a first community contribution window closing August 6.
The jobs reckoning intensifies
Monday.com said it will cut roughly 630 roles — about 20% of its workforce — to fund a shift toward its AI Work Platform, taking a $45–55 million charge in the process. The platform pairs a no-code app builder with customizable agents and a chatbot that generates reports and updates dashboards. The move is striking because Monday.com is not a struggling company; it is a profitable enterprise-software vendor restructuring its organization, not just its product, around AI. Amazon also confirmed layoffs within its artificial general intelligence unit even as it reaffirmed heavy AI investment. These are early data points in a pattern worth watching: the cuts are increasingly hitting healthy companies making a strategic bet, not distressed ones cutting to survive.
AI's content flood, quantified
Deezer disclosed that AI-generated tracks now account for more than half of daily uploads to its platform — about 90,000 fully synthetic songs a day as of June 2026. The streaming service, which says its detection tool runs at 99.8% accuracy, will systematically remove AI tracks tied to streaming fraud (up to 85% of streams from fully AI music were fraudulent in 2025) and those inactive for six-plus months. "Now that half of all daily uploads are AI-generated tracks, we are taking additional steps to safeguard the rights of artists and songwriters," said CEO Alexis Lanternier; the company is now licensing its detection tech to Billboard and Sacem. In parallel, Netflix told investors roughly 300 of its titles used generative AI this year. The provenance problem is no longer theoretical — it is a majority of the inputs on at least one major platform.
Quick Takes
Science Corporation's PRIMA vision chip won European regulatory approval to sell a device that restores functional vision lost to age-related macular degeneration via an implanted chip and camera glasses; it also holds a U.S. FDA breakthrough designation.
Travis Kalanick's robotics venture, Atoms, raised $1.7 billion led by a16z, with Uber joining — reconnecting the former Uber CEO with the company he founded.
Tesla will record employees at its Grünheide plant to generate training data for its Optimus humanoid robot, even as it pushed Cybercab, Semi, and Megapack 3 volume production past 2026.
U.S. Treasury kept sanctions "on the table" after the White House accused China's Moonshot of improperly distilling Anthropic's Fable model into Kimi K3; some experts dispute the distillation claim, noting Fable has only been public since July 1.
Google's Gemini 3.5 Pro remains delayed and in partner testing, months past its original June target, per Pichai on the earnings call.
Anthropic launched a Claude connector for its Economic Index, letting users query occupation, region, and automation-trend data in natural language.
What This Means for Your Business
The clearest operator takeaway this week is about cost architecture, not model choice. Cursor Router and OpenAI's own agent tooling both encode the same lesson: the frontier model belongs on the hard 10% of a task — planning, judgment, ambiguous decisions — while a cheaper, faster model handles the mechanical majority. If your AI bill is climbing, the fix is rarely a different vendor; it is routing work to the right tier. Audit where you are spending premium tokens on commodity work, and test a router or a two-model split before you assume you need to cut usage.
Second, the enterprise agent platforms are worth a real evaluation now. With Presence, Gemini Enterprise, Meta's platform, and Nvidia–ServiceNow all shipping governance-first products in a single month, the parts that used to require a bespoke engineering project — permissions, escalation, simulation, monitoring — are becoming off-the-shelf. For a smaller operator, that lowers the cost of a customer-support or internal-ops agent from a build to a configuration. The 75% resolution rate OpenAI cites on its own support line is a plausible near-term target, not marketing. But note the delivery model: Presence is white-glove and enterprise-priced, so smaller teams should watch for the self-service tier that competition will eventually force, and pilot on a narrow, well-bounded workflow first.
Third, treat the layoffs as a strategy signal, not just a labor story. Monday.com is telling the market it can grow revenue with fewer people by reorganizing around agents. Whether or not that bet pays off, your competitors are making the same calculation. The defensible response is not to cut headcount reflexively but to identify which repeatable, high-volume workflows in your business — reporting, reconciliation, first-line support, data entry — could be handled by a governed agent, and to build the muscle to deploy one safely before the pressure to do so becomes existential.
Finally, provenance is becoming a compliance and brand issue. Deezer's numbers and Netflix's disclosure show that AI-generated content is now a material share of real platforms, and detection, labeling, and fraud controls are following fast. If your business publishes content, sells on marketplaces, or licenses media, expect provenance requirements — and the ability to prove what a human made versus what a model generated — to move from optional to expected. Start documenting it now, while it is cheap to do voluntarily rather than expensive to retrofit under a mandate.