Alibaba's AI Finds Real Superconductors While Palantir Calls Tokens a Wealth Tax

So, What Actually Happened?

So, I went looking for the big model launch this week and it never showed up. What showed up instead was AI acting less like a chatbot and more like plumbing. We scanned 190,000 articles this week so you don't have to, and the same shape kept repeating. In China, Alibaba's agent found four new superconductors and had them verified in a lab, real physics, not a better paragraph. In the U.S., Palantir's CEO called token pricing a ”wealth tax” on enterprises. Quietly, down in the data trenches, the table-format war ended. And Sam Altman floated handing Washington a 5% stake in OpenAI. Different desks, same move: nobody argued about which model is smartest, everyone argued about what AI costs, what it produces, and who owns it.

The Bottom Line: The smartest model stopped being the story. This week AI got judged like infrastructure, on what it costs to run, what it builds, and who holds the keys.

The Tracks That Matter

1. Alibaba's AI Agent Just Discovered Four Real Superconductors

Every AI demo this year was a chatbot getting better at talking. This one is different. Alibaba's Damo Academy built an agent called Elements Claw that discovered four previously unknown superconductors, then had them confirmed in physical lab experiments. To put that in scale: the standard SuperCon reference database, built over decades, lists only about 2,000 of them. The agent screened 2.4 million candidate crystal structures in 28 hours of compute and narrowed 68,000 possibilities down to a handful worth testing. Here is the shift nobody's pricing in: as Alibaba leans hard into AI for the back half of 2026, the interesting AI stopped summarizing what humans already wrote and started producing knowledge humans didn't have. That's a different product category, and a different moat.

Here's what works: Stop asking AI vendors what their model can say. Ask what it can find, discover, or generate that isn't already in the training data. Output beats eloquence.

2. Palantir's CEO Calls Token Pricing a ”Wealth Tax”

While everyone debated model quality, Palantir's Alex Karp went on TV and reframed the entire AI cost conversation: enterprises, he argued, are effectively paying a ”wealth tax” on tokens to the frontier labs, metered by the word, forever. The market liked the framing; Palantir rallied 3.7% and DA Davidson upgraded it the same day. He is pointing at something real. A parallel look at AI token economics lays out the trap: when your AI is billed per token, every product you ship carries a variable cost that scales with success. The more customers love it, the more you owe. That is not a software margin, that is a utility bill wearing a SaaS logo.

Here's what works: Before you scale any AI feature, model its token cost at 10x usage. If the margin inverts as adoption grows, you have a pricing problem, not a product.

3. The Data-Lake Format War Just Quietly Ended

No press conference, no keynote, but for anyone running a data platform this is the week that mattered. A widely shared teardown declared the table-format war over: after years of Iceberg versus Delta versus Hudi, the industry has effectively standardized on open table formats, and the fighting is done. This is the boring, load-bearing stuff, the plumbing under every AI workload, and standardization here is worth more than another model release. It also explains why Databricks and Snowflake skill sets stopped overlapping: once the format is settled, the competition moves up to the engine and the governance layer on top. For once, the foundation got decided before the skyscraper leaned over.

Here's what works: If you're still hedging across three table formats, stop. The war is called. Pick the open standard, consolidate, and pour the saved effort into the governance layer where the real fight moved.

Quick hits:

• Washington might buy a piece of the AI boom. Sam Altman floated giving the U.S. government a 5% stake in OpenAI, modeled on Alaska's oil fund, which would put roughly $42.6 billion of public equity into a single lab and a very live debate into Congress.
• The market crowned the AI defenders. As reports spread that AI can now find software vulnerabilities, cybersecurity stocks hit records, with Palo Alto Networks up 14.5% on the week, a tell that AI's next winners may be the ones cleaning up after it.
• China's running the marathon, not the sprint. An investment note argues China is late to the AI sprint but strong in the marathon, and Alibaba's superconductor agent is exactly the kind of patient, full-stack bet that thesis predicts.

Signal vs. Noise

🟢 Signal: Automation. The quiet winner this week wasn't a model, it was automation itself, climbing in real influence while the hype cooled around it. That's the tell that budgets are shifting from what AI says to what AI actually does on the floor, and most coverage is still chasing launch announcements instead of the workflows getting rebuilt underneath them.

🔴 Noise: ”AI governance” as a headline. The governance label pulled some of the heaviest mention volume again this week, but its real pull faded as the concrete action moved to what AI produces and what it costs to run. Governance still matters enormously; tracking the word itself as this week's story just means reading last quarter's frame.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

Alibaba's agent discovered real superconductors, Palantir's CEO branded token pricing a ”wealth tax,” and the data-lake table-format war quietly ended: three stories from three completely different desks that are secretly the same story.

The tech press files the superconductors as a research curiosity. The markets desk writes up Karp's comment as a stock mover. The data-engineering blogs treat the format war as inside baseball. Read them on one page and the pattern is hard to miss: this was the week AI stopped being graded as a conversation and started being graded as infrastructure. Infrastructure gets judged on three things a chatbot never had to answer for: what it produces (real materials, not summaries), what it costs to run (metered tokens, not a flat fee), and what foundation it sits on (a settled data layer, not a format war). The move on Monday is to re-score your own AI initiatives on those three axes instead of on demo quality. The projects that produce something, control their run-cost, and sit on governed data are the ones that survive contact with a budget review. The rest are just very expensive conversations.

By The Numbers

• Alibaba's AI screened 2.4 million crystal structures in 28 hours and found four new superconductors: the entire SuperCon reference database, built over decades, holds only about 2,000. AI just added to the canon of physics, not the chat log.
• A 5% U.S. government stake in OpenAI would be worth roughly $42.6 billion: against the company's reported $852 billion valuation. That's the price tag on nationalizing a slice of the AI boom.
• The AI theme now accounts for about one-third of the net income of all S&P 500 companies: per Wellington. When a third of corporate America's profit rides on one theme, it stops being a sector and becomes the weather.
• The Russell 2000 surged nearly 22% in the first half of 2026, its best start since 1991: the AI rally finally spilled out of the megacaps and into small caps.
• Palo Alto Networks jumped 14.5% and CrowdStrike 10.7% in a single week, both to all-time highs: as reports spread that AI can now find code vulnerabilities, the market bet on the defenders.
• See what's rising in our 190K-article corpus this week →

Deep Dive: AI's Per-Stream Moment

I lived through one metering revolution already. When streaming replaced vinyl and CDs, the music didn't change, the billing did. Artists went from selling an album once to earning fractions of a cent per play, and they learned a brutal lesson: whoever gets metered per unit carries all the risk. AI is having that exact moment right now, and most companies are the artist who hasn't opened the statement yet.

The meter is always running
Token pricing is per-stream pricing. Every prompt is a play, and the lab takes its cut on each one. Karp's ”wealth tax” line landed because it's accurate: when you're billed by the token, your costs climb with your success. Love your product to death, and the meter loves you right back, straight out of your margin.

The track is getting cheaper, the catalog bigger
Meanwhile the models themselves are commoditizing. Open weights, Chinese full-stack players, a new capable model every few weeks. The ”track” everyone fought over is turning into a jukebox where every song costs about the same. Paying a premium meter for a commodity play is how budgets quietly bleed out.

Own your masters
Here's the lesson the smartest musicians learned: own your masters. The artists who kept their recordings kept their leverage. In AI, your masters are your governed data and your data foundation, the one thing the meter can't tax and the model can't commoditize. Alibaba's superconductor bet works because they own the full stack, not because they rent the cleverest model.

What Actually Works

1. Price AI like a utility, not a miracle: Model token cost per workflow at 10x scale before you commit. If margin inverts as you grow, fix it now, not at renewal.
2. Own your masters: Your governed data is the asset that survives every model swap and every price hike. Fund it like it's the product, because it is.
3. Treat the model as rentable: With open and full-stack challengers everywhere, don't build a moat on a track anyone can license next quarter.
4. Measure output, not eloquence: Alibaba's agent found real superconductors. Grade your AI on what it produces, not how well it talks.

The music never paid the artists who assumed the platform had their back. AI won't either. Read the statement, and own your masters.

What's Coming

AI-Discovered Science Goes Mainstream

Alibaba's superconductor find seeds a wave of ”our AI discovered X” announcements across materials, chemistry, and drug discovery. The smart move is to watch which ones get independently reproduced. That's where the real signal separates from the press release.

The Token-Pricing Revolt Spreads

Karp's ”wealth tax” framing is the opening shot. Watch enterprises push back on per-token metering the way they once fought per-seat overages. Flat and platform pricing will become a sales weapon, and the labs will have to answer for the meter.

Government Equity in AI Becomes a Real Fight

Altman's 5% stake proposal sounds like a trial balloon, but it has precedent now that Washington already holds a stake in Intel. Expect the sovereign-AI-fund idea to move from op-ed to committee hearing before year-end.

For Your Team

Strategic purpose: This week moved the AI question from ”which model is smartest” to ”what does ours cost to run, what does it produce, and who owns the foundation.” The teams still shopping for models are about to get schooled by the teams reading their token bills.

Monday's meeting prompt: ”If every AI feature we ship is metered per token, does anyone here actually know what our AI bill looks like at 10x usage, and what happens to our margin when we get there?”

Share-worthy stat: The AI theme now accounts for roughly one-third of the net income of all S&P 500 companies. A third of corporate America's profit is riding on one bet.

Go deeper: Track where AI value is really shifting →

The Track of the Day

”Enterprises are effectively paying a 'wealth tax' on tokens.”
— Alex Karp, Palantir CEO, on frontier-lab pricing

The model was never going to be the expensive part. The meter is. Read the statement before you sign.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: July 4, 2026 | Curated by Yves Mulkers @ Ins7ghts