On Tuesday, three AI chip companies raised $2.55 billion in one day. The inference wars are moving from research to capital.

Three artificial intelligence chip companies raised a combined $2.55 billion on a single Tuesday last week, each claiming to challenge Nvidia in some portion of the inference stack. SambaNova Systems closed the first tranche of a $1 billion Series F at an $11 billion valuation. Positron, the Reno-based chip startup, entered talks to raise $750 million at up to $5 billion. And China's Iluvatar CoreX completed an $800 million-plus share sale on the Hong Kong exchange at a 15% discount to close. Same day, same target, different geographies and different architectures.

This convergence was not planned coordination. It reflects something simpler: when the leading inference chip companies all surface simultaneously, it signals that the capital window for betting against Nvidia's inference monopoly has cracked open and investors are moving. The week of July 6 also brought $300 million into the vibe-coding platform Lovable at a $13.2 billion valuation, a $100 million raise by an AI agent startup whose agent did the actual fundraising, and the first real funding round for Ollama, the open-source tool that 8.9 million developers already use to run models locally. The week produced $17.9 billion across 56 rounds. Strip Blue Origin's $10 billion debut VC raise and you get $7.9 billion in core technology rounds, up from an adjusted $7.2 billion the prior week.

The AI infrastructure analysis is the obvious read. The less obvious one: this week, for the first time, the most interesting story in AI startup funding was about proof rather than promise. Lyzr AI let its own agent handle investor communications for a $100 million raise. Ollama raised money after demonstrating distribution at a scale most funded startups never reach. And post-quantum cryptography, after years of NIST standards committee work, attracted over $1 billion in commercial investment. The market is no longer purely buying potential. Some of it is buying demonstrated outcomes.

The numbers

The median check jump is the most durable signal. Headline totals are distorted by single-round outliers in any given week, but a $3.3 million increase in median check size reflects broad upward pressure across the distribution, not just the top end.

On one Tuesday, three Nvidia challengers raised $2.55 billion

SambaNova Systems completed the first close of its $1 billion Series F round, led by General Atlantic, with participation from BlackRock, Intel Capital, Qatar Investment Authority, T. Rowe Price, and Vista Equity Partners, among others. The company was valued at $11 billion. CEO Rodrigo Liang told TechCrunch that a second close was weeks away. On the customer side, JPMorganChase selected SambaNova as its inference infrastructure partner, deploying SN40 and SN50 systems for on-premises AI inference. That combination of institutional capital and enterprise contract is the format that has preceded several AI infrastructure IPOs in the past 18 months.

Positron, based in Reno, entered talks to raise $750 million in two tranches, with the first valuing the company at $3.5 billion and the second at up to $5 billion. The company's current shipping product, the Atlas chip, claims more than 3.5 times the performance-per-watt of Nvidia's comparable inference offerings in certain workloads. Its next chip, Asimov, is targeting tape-out in late 2026. Positron has already raised $230 million, including investment from Qatar Investment Authority and Arm. The structure of the raise is interesting: a two-tranche round with escalating valuations is typically used when a company wants to let technical milestones unlock better terms, which suggests Asimov's progress will determine which valuation they actually close at.

Iluvatar CoreX completed a share sale that raised approximately $900 million, discounted 15% to its closing price on the Hong Kong Stock Exchange. The company's shares had risen 234% since its January IPO, reflecting demand for non-US alternatives to Nvidia's GPU stack. Iluvatar is reportedly in talks to supply ByteDance with at least 50,000 chips in 2026, most for inference. That customer relationship matters: ByteDance is one of the few companies at scale that can absorb inference hardware at that volume and validate real-world performance claims.

The coordination here was market timing, not strategy. When SambaNova announced at $11 billion, it reset the comparison point for every other AI chip company in fundraising discussions. Positron and Iluvatar's teams, like any rational actors, moved faster on their own timelines once the benchmark valuation was public. The week effectively established SambaNova as the current floor for how the market prices credible Nvidia alternatives in inference. Both Positron and Iluvatar are priced well below that, which either reflects realistic risk discounting or leaves upside for their next rounds.

Agentic AI's most interesting proof of concept cost $100 million

Lyzr AI, a three-year-old enterprise AI agent platform based in Jersey City, raised $100 million at a $500 million valuation. The story everyone will remember from this round is how they did it: the company's own agent, SivaClaw, handled investor communications, fielded questions from more than 130 investors, drafted investment memos, tracked which deck slides investors spent the most time on, and generated $400 million in stated investor interest without the founders doing a single in-person Sand Hill Road meeting. Humans made the final decisions, but the agent ran point on everything before those decisions.

This is a demonstration that will be replicated. Every enterprise AI agent company now has a strong incentive to let their agent handle at least part of the next fundraise, not because it is more efficient (it may not be), but because it is the most credible possible product proof. The investor who reads that an AI agent startup used its own agent to raise money is the investor who has already seen the product in a high-stakes context. Norm Ai raised $120 million at a $1.2 billion valuation the same week, building AI agents for legal and regulatory compliance. Lovable, whose entire product is an agentic interface for building software through natural language, entered talks to raise $300 million at $13.2 billion, exactly double the $6.6 billion valuation it achieved in December. Lovable's reported annualized revenue run rate is $500 million.

The language shift is quantifiable. Among the 200 new startups published to our database this week, the term "agentic" appeared 95 times in descriptions, and "agent" 177 times. For comparison, the term "RAG" appeared 63 times and "foundation model" 64 times, both of which were the dominant technical descriptors two quarters ago. This is not just a labeling trend. Startups are structuring their products around agent orchestration in ways that require a different go-to-market motion than SaaS. The Lyzr fundraise is the clearest signal yet that founders have internalized the implication: if your product is an agent, the most powerful sales asset is showing the agent doing real work in a real-stakes context.

Ollama raised $65 million after building a moat most funded startups never reach

Ollama, the open-source tool for running language models locally, raised a $65 million Series B led by Theory Ventures. The company has 8.9 million monthly developer users, sits in 85% of the Fortune 500, and operates with 14 employees. In total it has raised $88 million.

The conventional VC model optimizes for funding-then-distribution. Ollama did the opposite. It launched in 2023, grew to nearly 9 million users without raising, and only now is raising institutional capital. The Series B values the company at a level that reflects the distribution, not just the product, which is how it should work. Theory Ventures, which led this round, has made a pattern of backing developer tools with large organic user bases. The signal from this deal is specific: community-first distribution for technical tools is now a recognized and priceable moat, not just a nice-to-have before real fundraising.

The business model question is unresolved. Ollama is open source, which means the path from 8.9 million users to revenue is not as direct as a SaaS product. But the Fortune 500 penetration suggests enterprise use cases are already there, and the capital will likely go toward building the managed or commercial tier that converts that enterprise footprint into recurring revenue. The 14-employee constraint also signals something: Theory Ventures is betting that the team can build commercial infrastructure without needing to scale headcount proportionally to users, which is only possible if the product is genuinely self-serve at scale.

Post-quantum security stopped being theoretical this week

Keyfactor, which manages encryption keys and certificates across enterprise infrastructure, raised $1 billion, the largest pure cybersecurity round of the week. QIZ Security raised $17 million at seed for a post-quantum readiness platform.

The US National Institute of Standards and Technology finalized its post-quantum cryptographic standards in 2024. Since then, regulated industries have been under informal and, increasingly, formal pressure to document their quantum readiness. The Keyfactor raise is a bet that enterprise migration from classical to post-quantum encryption will drive years of managed certificate infrastructure work. The QIZ raise is a bet on the audit and assessment layer that precedes that migration. Both bets can be correct simultaneously: the assessment market tends to precede the remediation market by 12 to 24 months in enterprise security cycles.

What is notable about both rounds is the timing. Two years ago, post-quantum security was primarily a government and defense procurement story. This week it attracted institutional capital into commercial enterprise software. The NIST finalization created the compliance hook that makes budget allocation politically safe inside large organizations. Companies that started post-quantum programs in 2023 are now in the middle of their deployment cycles, and that creates a visible procurement signal for investors. The $1 billion Keyfactor round is the first time that signal has translated into a mega-round in this specific subcategory.

Valuation velocity: Lovable and Mercor are both doubling in under a year

Lovable's $13.2 billion valuation, if the round closes, will represent a doubling from its December 2025 valuation of $6.6 billion in roughly seven months. Mercor, the AI training data platform, is in discussions to raise $500 million at a $20 billion valuation, which would be double its $10 billion October 2025 Series C valuation in nine months. Mercor's CEO announced an annualized revenue run rate above $2 billion, a 100% increase from four months prior.

Two 2x valuations in under a year, for companies in different segments (vibe coding and AI training data), reflect the same underlying pressure: the market is trying to price AI-native companies that have escaped the S-curve constraint of traditional software. Lovable is growing on consumer adoption by non-developers. Mercor is growing on structural demand from AI labs that need specialized human-generated training data. Neither company has a natural ceiling that the market can model clearly, and when ceiling uncertainty is high and revenue growth is fast, valuations overshoot. Whether both valuations are sustainable depends on whether the underlying revenue trajectories hold through 2026. At current growth rates, both are buying time to answer that question.

Microtrends worth watching

• World model data. Worldmodeldata raised £7 million (approximately $9 million) to convert gaming data into AI training sets for world models. This is the first specialized data company we have tracked that is explicitly targeting world models rather than language models or image models. If world models become a serious research priority at the major labs in the next 12 months, this category will expand quickly.
• Quantum computing with a falsifiable claim. Oratomic raised $300 million for a quantum computer design it claims requires only 20,000 qubits to be commercially viable. The technical claim is specific enough to be falsifiable, which makes this a more interesting bet than most quantum rounds, which paper over qubit counts with general market potential language.
• European fusion gets its anchor company. Proxima Fusion raised €411 million at a €2.4 billion valuation, positioning itself as Europe's commercial fusion anchor. The US has had Commonwealth Fusion for several years. Europe getting a comparably capitalized contender has geopolitical implications for clean energy independence that extend beyond startup funding dynamics.
• Open-source AI infrastructure gets institutional validation. Prime Intellect raised $130 million for its distributed AI training platform. Between Prime Intellect and Ollama, open-source AI infrastructure attracted roughly $195 million this week, which is meaningful given how hard it has historically been to raise institutional capital for open-licensed products.
• Figma is collecting agent teams. Figma acqui-hired the Bud team, a YC-backed agent platform that could browse the web, write code, and automate tasks. Bud is shutting down on July 18. This is the second AI agent acqui-hire for Figma in two quarters. A pattern of two is not proof of intent, but it is a signal worth watching ahead of their next product announcements.

What might happen next week

Two predictions, both specific enough to be wrong.

First: Positron will announce a first close on its $750 million round at the $3.5 billion tranche valuation, not the $5 billion target, before the end of July. The two-tranche structure gives the company a rational reason to close the first tranche quickly while momentum from the SambaNova announcement is still fresh. The $3.5 billion valuation is the more defensible anchor given Asimov has not yet taped out. The $5 billion will wait for the technical milestone.

Second: a second post-quantum security company will raise a round above $50 million within the next two weeks. The Keyfactor and QIZ Security closes this week will create a reference point that makes it easier for similar companies to close on favorable terms. Compliance-driven enterprise security tends to move in clusters once the first institutional check lands. The Keyfactor round is that check.