The Investor

From fees to firm-level penalties

Last year the administration raised H-1B fees. This year it is punishing specific companies, The Information reports. Microsoft's H-1B workers can no longer apply for permanent residency through the company, and universities' J-visa sponsorship is under investigation. The stated trigger was a claim that Microsoft laid off Americans while hiring immigrants. Vance said each laid-off worker was replaced by "one and a half foreign indentured servants." The announcement came on the same day Satya Nadella was due to receive a National Medal of Technology at the White House. No company is politically insulated.

Microsoft's rebuttal also tells you who is exposed. It says about 4,800 of its roughly 6,000 filings were for people it already employed in the US, which Morning Brew puts at about 1% of its US workforce. The people affected are incumbent employees partway through the green-card process, not future hires. That makes this a retention shock as much as a hiring one. The Information names "Tata Communications." The large H-1B sponsor in this debate is usually Tata Consultancy Services, so confirm which entity is meant before acting.

The labor-arbitrage floor was already cracking

The firms most exposed are the Indian IT outsourcers. Techpresso notes that nearly three-quarters of H-1B visas go to workers from India. a16z's Revelio data shows global call-center employment turning negative in every income tier by Q4 2025. High-income markets went negative first, in 2022, and the poorest markets went last. Policy and automation are now squeezing the same business model from two sides. The outsourcers' enterprise clients also have a compliance reason to re-bid work.

Who actually wins the re-bid

Some sources call this an outright tailwind for AI-native services. The reliability data argues for more caution. Techpresso reports that top models failed to rederive nearly half of classic algorithms once memorization was blocked. In our view, the contracts that move first will go to vendors that sell outcomes and keep human approval steps, not to fully autonomous replacements. A second opening is more immediate. The Information suggests startups able to sponsor visas can recruit Microsoft engineers whose green-card path has stalled, with immigration counsel managing transfers.

The Information flags one more pattern to screen for: recent layoffs combined with continued visa hiring is the combination that triggered enforcement. Any portfolio company showing both carries headline risk. That risk is now enforced against named employers, not just debated.

Visa policy has moved from a cost line to an enforcement risk aimed at specific employers, and the labor-arbitrage model is the first business it reprices.

Why free is worse than cheaper

Microsoft ran this playbook against Slack with Teams. The suite owner prices the add-on at zero and collects on seat upgrades it already sells. The current version is broader. AI Breakfast reports that Google's Gemini agent ships with attested agent identity, an Agent Sandbox, an Agent Gateway, Smart Routing and hard per-project spend caps. Each one maps onto a category venture investors have been funding: non-human identity, AI gateways, model routers, agent sandboxes and AI FinOps (software for controlling AI spend). Google is shipping all five into a base where nearly 90% of the Fortune 100 already use Gemini Enterprise.

The same thing happened further down the stack. TLDR Hardware reports that Google open-sourced ML Drift, a GPU inference runtime that works across chip platforms and was co-optimized with Arm, Intel and Qualcomm. AWS released an open-source Physical AI toolchain built on Nvidia's Isaac, GR00T and Cosmos stack. In security, CyberScoop reports that CrowdStrike, Palo Alto Networks, Booz Allen and Rockwell Automation chose to partner with Anthropic rather than build or buy. Read plainly, they judged building their own not worth the budget.

Where each bundle stops

The Batch reports that an unnamed Gemini model hacked into three other companies' systems during autonomous testing because of poor sandboxing, which Google and security firm Irregular confirmed. The Instagram AI-support account takeovers, reported by 404 Media, needed no prompt injection at all. The support agent took a code verified at a mailbox as proof that the person owned the account. A mailbox proves rather less than that. Both failures sat at authorization and containment, a layer a single-suite control cannot fully cover for an enterprise running agents across several clouds.

Where the reads diverge

TLDR's read is that the categories are validated and the room for startups has shrunk. AI Breakfast raises the more interesting puzzle, which is pricing. If Google gives each agent its own directory account, seat-based SaaS gains a new unit to bill. If Google declines to charge per agent, every per-seat vendor faces pressure to move to usage pricing. Google's eventual Gemini agent price will set the comparison for the whole category. Google has not published that price, so investors are valuing these companies against a comparison point that does not exist yet.

The buyer who pays a startup for neutrality is the enterprise running agents across several clouds, because a free Google control is built for Google's own suite.

Our view, which could be wrong on timing, is that a horizontal agent company whose customers sit mostly on Google Workspace is the most exposed. A company selling authorization, outcome validation or payment guardrails across clouds is the least exposed, and the bundles raise its value by putting more agents into production.

Two deals, one test

Firmus asked public buyers to pay nearly triple its August private price for an operator with two data centers and no record of building at scale, The Information reports. Its strongest proof point was Nvidia's June agreement to back a 170,000-chip order through its AI Compute Partnership, a program in which Nvidia supports partners' chip purchases in exchange for a revenue share. Sources describe the outcome differently. Bloomberg says the IPO is set to be postponed and a private round is being weighed. The Information says the deal was pulled, citing volatility. Either way, the step-up did not clear.

Waymo's loan passed the test Firmus failed. It is Waymo's first debt ever, it was upsized, it is secured only by IP and it carries no Alphabet guarantee. Goldman arranged it. PIMCO, Blackstone and Sixth Street led, and Apollo, Blue Owl, HPS and Oaktree joined. The loan is about 4% of the $126B valuation set in Waymo's $16B equity raise, and the service already runs in 15 US cities. The lenders were paying for proof, and they had a large cushion behind it.

Demand is not the variable

Chip demand keeps compounding. TLDR Hardware reports TSMC's September revenue at NT$511.86B, up 54.6% year over year, and Bloomberg puts quarterly growth at 51%. The friction sits below the chips. Bloomberg calls Oracle's gas trucking "expensive" and does not link it to the share drop. It reflects a power bottleneck, not a financing signal: the pipelines are late, not the money. Still, the choice tells you something. A hyperscaler would rather pay a bridge-fuel premium than accept delay. That makes bridge and behind-the-meter power a category with proven willingness to pay.

Anduril shows the same pattern in defense manufacturing. It is committing $3.7B of capex to a shipyard against a Navy contract worth up to $2.9B, paid only on demonstrated output. Production starts in California in 2028 and Maryland in 2030. The outlet reporting it is a disclosed Anduril investor. Physical bottlenecks are being funded by private balance sheets large enough to carry the risk until the asset proves itself.

What it means for your book

When an issuer switches from an IPO to a private raise, leverage on price and structure moves to private capital. The first diligence question for Firmus is still unanswered. The Information reports that Nvidia paused some partnership deals in August and is rewriting contracts, but it does not say whether Firmus's arrangement is affected. Waymo's structure is a template only for companies with comparable scale. Applying it to subscale companies would misread why the lenders said yes.

A near-tripled step-up for an unproven operator did not clear public markets, but private lenders will fund IP alone when the business already works.