Big 5 AI Vendor Roundup: Week of July 13, 2026

Technology Note By: Mark Tauschek, Bill Wong, Info-Tech Research Group

This week the enterprise AI fight moved to the plumbing. Google, Microsoft, Salesforce, Snowflake, and ServiceNow lined up behind a shared standard for connecting AI agents to business software, a move being framed as a direct push against Anthropic and OpenAI. The target is Anthropic's Model Context Protocol, which quietly became the default over the past 18 months, and the five companies that hold most of the world's business data would rather not build the agent era on a competitor's foundation. Meanwhile Google's own flagship problem got worse: Gemini 3.5 Pro has now missed its June target by a month, reportedly after a ground-up rebuild, knocking a few percent off Alphabet's stock. OpenAI shipped a full-duplex GPT-Live voice model and got sued by Apple. Anthropic pushed harder on state AI laws. AWS had a quiet week.

The theme is a step beyond last week's deployment story. The models are close enough in capability that the fight has moved to who owns the standard agents run on, who can ship a flagship at all, and who sets the rules. For IT leaders, the primary consideration from this week is vendor lock-in.

The agent standards war breaks into the open

For most of the past two years the interesting AI questions were about models. This week the most consequential one was about a protocol.

  • Google, Microsoft, and a group of enterprise software vendors backed a rival agent standard. The Information reported Wednesday that Google, Microsoft, Salesforce, Snowflake, ServiceNow, and others agreed to support a shared standard for connecting AI agents to business software, in a move it framed as an effort to beat back Anthropic and OpenAI. The target is Anthropic's Model Context Protocol (MCP), which Anthropic open sourced in late 2024, donated to the Linux Foundation's Agentic AI Foundation in December 2025, and which now sits at roughly 97 million monthly SDK downloads with adoption across every major lab. MCP has effectively won the agent to tool layer, and that is the problem for the companies whose software holds most enterprise data. They would rather not have the connective tissue of the agent era defined by a rival. According to Crypto Briefing, the new specification is called Agent Resource Discovery (ARD), a federated discovery layer under Apache 2.0 licensing with additional backers including Cisco, Databricks, GitHub, NVIDIA, and Hugging Face, and it is positioned to sit alongside MCP rather than replace it outright. Neither OpenAI nor Anthropic is on the initial supporter list.
  • Why this matters more than a model release. Standards wars decide who owns a technology era, and this one is being fought over invisible infrastructure that will be very hard to rip out later. If you are building agent workflows now, you are implicitly betting on a standard. The pragmatic move is to treat the plumbing the way you treat any integration layer: prefer open, vendor-neutral standards where you can keep your agent-to-tool connections abstracted rather than wired to one vendor's proprietary format, and watch whether the new alliance produces something genuinely interoperable or just a walled garden with better marketing. MCP's Linux Foundation governance is a point in its favor here, and the alliance has not yet shown its hand on openness.

Google: The flagship rebuild, and coleading the standards fight

Google had the heaviest and least comfortable week of any vendor.

  • Gemini 3.5 Pro has slipped by another month after a reported rebuild. The flagship Sundar Pichai promised "next month" at May's I/O has now missed that target by a wide margin, and Bloomberg reported Google is taking extra time to improve the model's coding after a late June retraining attempt produced disappointing results. Multiple outlets report the delay is not routine polish but something closer to a ground-up pretraining restart, with the model falling short of Google's own internal benchmarks and trailing GPT-5.6 and Fable 5 on coding. Alphabet stock fell roughly 3-4% on the reporting. This lands on top of the senior researcher departures we covered in prior weeks. For IT leaders the guidance remains to not put Gemini 3.5 Pro on any 2026 roadmap you actually depend on. Evaluate Gemini 3.5 Flash, which is shipping and performs well, and treat Pro as upside whenever it arrives.
  • Google is coleading the standards alliance. As covered above, Google is one of the anchors of the new agent protocol push. It is a reminder that Google's platform strength – Workspace, Cloud, Android, and the enterprise software relationships around them – remains real even in a week when its model story went sideways. The company is choosing to compete on the layer it controls rather than the one where it is behind.
  • Intel picked Google Cloud for its own AI transformation. Intel and Google Cloud expanded a multiyear deal that puts Gemini Enterprise and the Gemini Enterprise Agent Platform across Intel's workforce, with custom agentic workflows for engineering, supply chain, and corporate functions and Google Cloud C4 and N4 instances to run chip design simulations. It is a useful counterweight to the flagship stumble because even with the model delayed, Google is winning large enterprise agent deployments on the strength of its platform, and landing a marquee semiconductor customer is a real proof point for Gemini Enterprise.
  • A quieter branding move. Google rebranded NotebookLM as Gemini Notebook and added a secure environment for running code, stretching the Gemini name across more products even as the flagship model slips. Useful for research and analysis teams already in the Google ecosystem but not a reason to change course.

Anthropic: A Canada research commitment, Meta compute talks, state law push, and a pricing stumble

  • CA$10 million for Canadian AI research. Anthropic committed CA$10 million to Canadian research institutions, partnering with the three national AI institutes (Amii in Edmonton, Mila in Montreal, and the Vector Institute in Toronto) plus hospitals and universities, largely through Claude credits. Alongside it, Anthropic published a Canadian country brief showing Canada ranks eighth worldwide in Claude.ai use and second by per-capita adoption. It is a modest sum, but it extends the sovereignty and research positioning Anthropic has leaned on since the export control episode, and it pairs with the Alberta government case study from the prior week, where Claude Code reviewed 466 million lines of provincial code in roughly 20 hours.
  • Reported talks with Meta for a $10 billion compute lease. Anthropic and Meta are in early talks over a roughly $10 billion deal for Anthropic to lease Meta compute capacity, according to reporting the two companies have declined to comment on. Treat it as unconfirmed and early. If it holds up, it is notable on two fronts: it would add yet another supplier to Anthropic's already unusual spread of compute relationships, and it would put Anthropic on Meta infrastructure even as the two compete on models, another instance of the circular compute economics running through this whole cycle.
  • Anthropic is spending to defend state AI laws. Anthropic is pursuing a state-by-state push for tougher AI safety rules, backing measures like California's SB 53 and New York's RAISE Act and funding midterm races through Public First Action, in direct contrast with OpenAI's preference for uniform federal rules that would pre-empt state law. This is worth watching not for the politics but for the compliance consequences. If Anthropic's approach wins, enterprises will face a patchwork of state AI obligations. If OpenAI's federal pre-emption approach wins, a single national standard. Either way, the regulatory map your AI governance program must satisfy is being drawn right now, and the two biggest labs are on opposite sides.
  • A pricing stumble worth noting. Anthropic briefly tested restricting Claude Code to its $100 per month Max tier and above, removing it from the $20 Pro plan, then reverted within hours after backlash, calling it a limited test affecting about 2% of new signups. Minor on its own, but a reminder that the economics of agentic coding tools are still unsettled and pricing can move quickly.

OpenAI: GPT-Live voice, and an Apple lawsuit

  • GPT-Live brings full duplex voice. OpenAI launched a new generation of voice models called GPT-Live that can listen and speak at the same time, which OpenAI says makes talking to the model feel much more like a real conversation than the old turn-taking approach. The strategic read is that voice is the interface where OpenAI's consumer lead is strongest and where Google, with Gemini built into Android, is its most dangerous long-term rival. Shipping GPT-Live days before Gemini 3.5 Pro was expected is OpenAI planting a flag on the one surface a Gemini price cut cannot easily take. For enterprises, voice agents are still early, but this raises the ceiling for customer service, field, and accessibility use cases worth piloting.
  • Apple sued OpenAI over trade secrets. Apple filed a trade secret suit against OpenAI on July 10, reportedly over hardware, and the fallout spilled into a public weekend argument between Elon Musk and Sam Altman on X. The specifics are unproven and the feud is mostly noise, but the signal is that OpenAI's ambitions now reach into hardware, where it will collide with Apple. Not an enterprise decision point today, but a sign of how wide OpenAI's fight is getting as it approaches its IPO.
  • The government precedent is now set. GPT-5.6 went fully public on July 9 after its government gated preview, and the White House pushed back on the idea that it gave any "green light," insisting release decisions rest with the companies. The formal voluntary framework is still pending, with an August 1 deadline for the multi-agency group to publish how the review process works. Until then, the de facto pattern from the Fable and GPT-5.6 episodes stands as frontier releases in the US now run through Washington, however it is labeled.

Microsoft: Anchoring the standards alliance, deepening MCP, and a shot at Anthropic

Microsoft's main move this week was co-anchoring the agent standards alliance, covered above. Three other things stood out.

  • Nadella took a public shot at Anthropic. Satya Nadella criticized Anthropic's Fable model for being "editorially controlled," an unusually direct jab from one lab leader at another. It fits Microsoft's broader positioning as the model-neutral platform, but it is also the kind of rhetoric that signals how competitive the enterprise fight has become. Treat it as marketing, not evaluation guidance, and test models on your own tasks.
  • The 3M partnership is Frontier Company in practice. Microsoft and 3M announced a strategic partnership with two parts. Microsoft becomes the first announced hyperscaler to deploy 3M's Expanded Beam Optical technology in Azure data centers, meant to speed fiber connections and cut maintenance in dense AI buildouts, and 3M becomes an early Frontier Company customer, with embedded Microsoft engineers to build an agent-driven workflow for 3M's order management (credit checks, delinquency assessments, system updates) under human-in-the-loop controls. It is a concrete look at the embedded engineering model from last week landing at a named customer and a reminder that the data center supply chain is now part of the AI story.
  • Copilot deepened its MCP support, the same week Microsoft backed a rival to it. Microsoft's July 15 Copilot release notes made MCP agents available inside Word, Excel, PowerPoint, Outlook, and Catalyst, added governed publishing to an enterprise Agent Store, and brought MCP-based federated connectors into the admin center. The irony is worth noting as Microsoft is leaning further into Anthropic's Model Context Protocol across its flagship productivity suite in the same week it co-anchored an alliance positioned against MCP. The reality for IT leaders is that MCP is not going anywhere at the tool layer regardless of the standards fight, and the more immediate item is governance, since the new Agent Store submission flow and federated connectors give admins a real control point for which agents and data sources employees can reach.

Amazon: A quiet week

AWS had a light week, notable mostly for who was not in the room.

  • AWS sat out the standards alliance. The new agent protocol push was led by Google, Microsoft, and the big enterprise software vendors, without AWS as an anchor. That fits AWS's positioning as the neutral, multimodel platform through Bedrock, where it hosts Anthropic, OpenAI, Google, and xAI models under one set of controls. Staying out of a standards fight aimed at two of its own Bedrock partners is consistent with that neutrality, and for IT leaders it keeps Bedrock a reasonable hedge while the protocol question shakes out.
  • The prior week's items are settling in. Fable 5 is back on Bedrock, the $1 billion Forward Deployed Engineering organization is live, and the Mechanical Turk wind down for new customers is still set for July 30. Nothing new landed this week that changes those.

Our Take

The story of the week is that the enterprise AI fight has moved from the model to the standard. When Google, Microsoft, Salesforce, Snowflake, and ServiceNow line up against Anthropic's MCP, they are fighting over the connective layer that will decide whose ecosystem enterprises are locked into for the next decade, and they are doing it precisely because the model race is too close to win outright. This is good and bad for buyers. Good, because a serious challenge to a single default protocol keeps the layer competitive and open. Bad, because a fragmented standards war means the agent connections you build this year could need rework depending on who wins. The defensive move is to keep your agent-to-tool integrations abstracted and to favor whatever stays genuinely open and vendor neutral, which today still favors MCP under Linux Foundation governance.

Google's week is the cautionary tale underneath all of it. A vendor that promised a flagship in June, missed it by at least a month, reportedly had to rebuild the model, watched its stock slide, and lost senior researchers is a reminder that roadmaps from even the strongest labs are intentions, not commitments. That is not a reason to write Google off, since its platform and distribution are still formidable and it is co-leading the standards fight from real strength. But it is a reason to build plans around what is shipping today, keep model choices portable, and treat every unreleased flagship, from any vendor, as a maybe.

What IT leaders should be doing:

  • Keep your agent integrations portable. With a standards war now open between MCP and a Google- and Microsoft-led alliance, abstract your agent-to-tool connections rather than hardwiring them to one vendor's format. Favor open, vendor-neutral standards until the fight resolves.
  • Do not depend on Gemini 3.5 Pro. It has slipped by another month after a reported rebuild. Evaluate Gemini 3.5 Flash, which is shipping, and treat Pro as upside with no committed date.
  • Pressure-test the new alliance before you commit to it. The Google- and Microsoft-led standard comes from the vendors that hold your business data, so the incentive to lock you in is real. Before adopting it over MCP, ask whether it will be governed openly, like MCP is under the Linux Foundation, or controlled by its backers.
  • Map your state versus federal AI compliance exposure. Anthropic is funding tougher state laws while OpenAI pushes federal pre-emption. Track both, because your governance program may need to satisfy a patchwork of state rules or a single federal standard depending on how this shakes out.
  • Pilot voice agents but measure them. GPT-Live raises the ceiling for conversational voice. Test it on real customer service or field workflows before committing, and watch consumption-based costs the same way you would for any agent.

Want to Know More?

Latest Technology Notes

All Technology Notes