Finswiss Labs · Insights

Australia: CPI vs House Price Inflation, 1997-2026

Finswiss Labs — Thu, 28 May 2026 00:00:00 GMT

The 48,000-Person Lever: How Samsung's Memory Workers Just Renegotiated the AI Supply Chain

Finswiss Labs — Thu, 21 May 2026 00:00:00 GMT

The 48,000-Person Lever: How Samsung's Memory Workers Just Renegotiated the AI Supply Chain

Finswiss Labs · What the Tech! · 21 May 2026

For about ninety-six hours this week, the bottleneck of the global AI build-out was not a hyperscaler's capex line or a Taiwanese fab. It was a bonus formula. On Tuesday evening (20 May 2026) in Suwon, lawyers for Samsung Electronics and the National Samsung Electronics Union (NSEU) sat across a table arguing about what fraction of operating profit belongs to the people physically running the world's largest memory plants. At 22:00 local time, with a strike scheduled to start in seventeen hours and JPMorgan estimating the daily revenue burn at roughly $700M, the company blinked. The 18-day walkout that would have idled 48,000 workers and disrupted Q3 HBM3E and HBM4 deliveries was suspended pending a ratification vote that begins tomorrow (22 May) and closes 27 May (CNBC, Bloomberg).

This is the story of a leverage discovery, not a labor dispute. The strike that didn't happen reveals something more interesting than the one that would have: how AI demand structurally changes the bargaining geometry of the firms that make the infrastructure underneath it.

The setup: a profit pool the union could see from the parking lot

Samsung's Device Solutions (DS) division, which contains memory and foundry, has been the part of the company most directly catching the AI updraft. The mass-production ramp of HBM4 began in February 2026 and the entire calendar-year production run is sold out before half-year (Fortune). HBM3E is the memory stacked next to Nvidia's H200 and Blackwell silicon; HBM4 is what gets soldered against Blackwell Ultra and Rubin starting in H2 this year. Every Hopper-class or Blackwell-class GPU sold by Nvidia at a ~75% gross margin (per Nvidia's Q1 FY27 print on the same day, also 20 May) carries a small stack of Samsung or SK Hynix output. The union noticed.

What broke the talks was not a wage number. It was the structural unfairness of a 50%-of-base-salary cap on performance pay during a period when the chip division was generating AI-grade returns. The union's opening position was 15% of operating profit allocated to a performance pool, no cap. Management's counter, before the collapse on 20 May, was 10%. The two sides settled at 10.5%, but the design of the payout is what makes the deal interesting, and it is the part most outlets are under-reporting.

The deal Samsung actually signed

The headline is that 10.5% of DS division operating profit is now earmarked for the bonus pool. The mechanism that makes the deal palatable to Samsung's CFO is a stock-vesting structure with a multi-year target gate:

Stock-funded portion vests over at least 10 years.
Vesting is contingent on the chip division hitting cumulative operating profit of KRW 200 trillion ($133.65B) over 2026 to 2028, then a lowered cumulative threshold of KRW 100 trillion across 2029 to 2035 (Bloomberg, TechTimes).

Translated into napkin math: if DS hits 200T won by FY28, the bonus pool generated by the 10.5% allocation totals roughly KRW 21T ($14B) over the three-year window. Spread across 45,000 to 48,000 union-eligible employees, that is a per-head ceiling on the order of $290k to $310k, deferred over the decade. That is not a salary, it is an equity-style retention instrument with operating-profit hurdles, and it is the kind of compensation structure Samsung has historically reserved for executives at the VP level and above.

The stock market read this correctly on Thursday morning in Seoul. Samsung Electronics common closed +6% on the news (CNBC). Investors decided the option-value of an averted 18-day production halt was worth materially more than the option-value of the deferred-stock liability the company just took on. Implied: institutional holders believe the AI capex cycle is structural enough to clear KRW 200T cumulative operating profit. That is itself a non-trivial market signal.

Why this was a workable threat in the first place

This is the psychology of leverage. Strikes work when the cost of disruption to the employer exceeds the cost of concession. Three numbers explain why Samsung was negotiating against an unusually steep cost-of-disruption curve:

$700M/day in lost revenue during the proposed 18-day action, per JPMorgan, with full-strike revenue impact above KRW 4T (Tom's Hardware coverage cited in daily brief).
2.1T to 3.5T won of operating-profit downside if the company conceded fully to the union's opening 15% ask, equivalent to a 7-12% hit to FY26 operating profit (JPMorgan).
An entire sold-out HBM4 production year, meaning that lost output cannot be recovered with overtime or excess inventory drawdown.

The third number is the load-bearing one. In a normal industrial-action environment a company can lean on inventory, accept some backlog, and recover after the action ends. In an environment where the product is allocated months in advance and the downstream customer is Nvidia, every disrupted shift is a missed delivery slot that propagates forward through Q3 GPU shipments, then forward again into hyperscaler training runs, then forward once more into the timing of model releases that pay for the GPUs. The chain has no slack. The union understood this; management's behaviour from the moment the public strike-vote landed indicates they understood it too.

Bank of Korea modelling cited by Al Jazeera projected 0.5 percentage points off Korean 2026 GDP in the full-strike scenario, an estimated KRW 30T (US$20B) macro hit (Al Jazeera). When a single bargaining unit at a single company can move a national GDP print by half a point, the bargaining unit is no longer playing the same game.

The cognitive trap the union avoided

There is a behavioural finance pattern that frequently undoes high-leverage labor actions: anchoring on absolute wage figures. Workers compare paychecks year over year, employers compare wage bills, both lose sight of the profit pool's growth rate. The NSEU's tactical insight, visible in the public framing of the dispute, was to anchor on the share of operating profit rather than the level of compensation. Once that becomes the metric, every quarter of Nvidia-tail AI revenue that flows through HBM strengthens the union's relative position automatically. Management cannot grow its way out of the demand. It can only renegotiate the share, which is precisely the conversation Samsung wanted to avoid for the entire decade prior to AI.

Compare this to Meta, which on the same day (20 May 2026) executed an 8,000-person involuntary RIF predicated on the substitutability of integrity, content-design and cybersecurity roles by AI workflows (Al Jazeera). The Meta workers had no analogous lever. They produced human-graded outputs into a stack where the marginal product had a credible AI substitute. The Samsung workers operate equipment with multi-billion-dollar replacement cost, multi-year qualification cycles, and zero credible AI substitute on a 12-month horizon. Both groups face the same AI capex cycle. One captured a multi-decade equity-style claim on it. The other got a 16-week severance offer.

What ratification week means

The deal is not done. The membership vote runs 22 May to 27 May. If members reject, the strike can be reactivated. Two considerations argue for approval. First, the structure converts wage-style demands into stock-style claims, which most union members are accustomed to seeing only on executive comp statements. Second, the 200T won three-year target is materially below the consensus DS division forecast for FY26 to FY28 inclusive, meaning the hurdle is plausibly clearable on current trajectory. The Samsung stock-price move on Thursday is the market's first vote on whether the hurdle is real or theatrical, and the answer it gave is theatrical.

The takeaway for Finswiss readers

The headline you will see in the wires this week is "Strike Averted." The headline that matters for anyone modelling AI supply chains is "Memory Labor Capitalises Itself." For thirty years the semiconductor industry has been a cost-pass-through business where labor was the residual claimant after capex and silicon. AI demand briefly inverted that hierarchy and the workforce closest to the constrained physical asset noticed before the market did. They formalised a multi-decade equity stake in the AI build-out while the equity holders were still distracted with Nvidia's earnings call.

The lever is now visible to every memory-fab union on the planet. SK Hynix's labor council should be reading the term sheet by Monday. So should the procurement teams at every hyperscaler.

References

CNBC, "Samsung Electronics shares rally 6% after union suspends strike following tentative wage deal," 21 May 2026. https://www.cnbc.com/2026/05/21/samsung-electronics-union-strike-suspended-wage-deal-bonuses.html
Bloomberg, "Samsung Union Suspends Strike, Reaches Tentative Deal," 20 May 2026. https://www.bloomberg.com/news/articles/2026-05-20/samsung-union-shelves-strike-plan-on-tentative-deal-yonhap-says
Al Jazeera, "Why are nearly 50,000 Samsung workers about to strike in South Korea?" 20 May 2026. https://www.aljazeera.com/news/2026/5/20/why-are-nearly-50000-samsung-workers-about-to-strike-in-south-korea
Al Jazeera, "Samsung Faces Chip Plant Strike That Threatens Global Supply" via Bloomberg, 20 May 2026. https://www.bloomberg.com/news/articles/2026-05-20/samsung-faces-risk-of-chip-disruption-after-labor-talks-collapse
Fortune, "A 45,000-person labor strike at Samsung's memory chip plants could throw a wrench into the AI boom," 17 May 2026. https://fortune.com/2026/05/17/labor-strike-samsung-ai-hbm-chips-dividend-revolution-memory/
TechTimes, "Samsung Electronics Averts 48,000-Worker Strike With Bonus Deferral Deal: Member Vote Runs May 22–27," 20 May 2026. https://www.techtimes.com/articles/316905/20260520/samsung-electronics-averts-48000-worker-strike-bonus-deferral-deal-member-vote-runs-may-2227.htm
Korea Herald, "[Breaking] Samsung Electronics union to launch strike after talks collapse," 20 May 2026. https://www.koreaherald.com/article/10741855
Washington Post, "Samsung's union puts off strike after reaching last-minute wage deal with management," 20 May 2026. https://www.washingtonpost.com/business/2026/05/20/korea-samsung-union-strike-memory/2b1eca4c-5403-11f1-9c40-7a0a12d9e745_story.html
Taipei Times, "Samsung reaches tentative union deal, averting crippling strike," 21 May 2026. https://www.taipeitimes.com/News/front/archives/2026/05/21/2003857701
Finswiss News Daily Briefing, 21 May 2026 (daily-tech-news-2026-05-21.md), Item #3.

Images and source credits

Samsung Hwaseong campus aerial (Bloomberg, photo credit SeongJoon Cho/Bloomberg). Linked at https://www.bloomberg.com/news/articles/2026-05-20/samsung-faces-risk-of-chip-disruption-after-labor-talks-collapse
NSEU union rally photograph (Al Jazeera, photo credit Yonhap/Reuters). Linked at https://www.aljazeera.com/news/2026/5/20/why-are-nearly-50000-samsung-workers-about-to-strike-in-south-korea
Samsung Electronics intra-day price chart, 21 May 2026 (CNBC market data). Linked at https://www.cnbc.com/2026/05/21/samsung-electronics-union-strike-suspended-wage-deal-bonuses.html
Hero image: Samsung signage, photo by Kote Puerto on Unsplash (free under the Unsplash License). https://unsplash.com/photos/samsung-signage-ILQoiHMJMME

Editor's note: image inclusion subject to source licensing; embed via canonical URL with on-page credit text matching the photographer name supplied by the originating publication.

Can AI be trusted? AI Mass Formation, when interacting in swarms.

Finswiss Labs — Mon, 18 May 2026 00:00:00 GMT

At Finswiss Labs we spend a lot of our time thinking about how AI behaves once it's out of the lab and inside a real business. Emergence AI just published a study on what happens when you leave swarms of autonomous agents alone in a simulation for a few weeks. And it's not good!

Emergence AI built a continuous-run simulation called Emergence World. A digital town with 40+ locations, real NYC weather, live news, and agents with assigned roles (scientist, explorer, mediator, engineer and so on). Each agent had to earn energy through action or die. Tools included navigation, voting, building, communication and, yes, arson.

They ran the same world several times. Same town, same roles, same rules. The only thing that changed was the foundation model powering the swarm of agents inside it. After 15 days, the five worlds looked nothing like each other.

Cumulative crimes by world over 15 days. Source: Emergence AI.

Five worlds, five outcomes. Source: Emergence AI.

Claude Sonnet 4.6 was the only model that survived with zero crimes. Gemini 3 Flash hit 683 crimes and was still climbing. Grok 4.1 Fast collapsed at 183 crimes by day 4. GPT-5 Mini's agents were law-abiding but forgot to feed themselves and were dead by day 7.

The headline numbers weren't the compelling part for us though.

In the mixed-model world, where Claude shared the town with Gemini, Grok and GPT, the Claude agents recorded crimes that they did not record when running alone. Emergence describes this as cross-contamination: an otherwise safe agent absorbing the norms of its peers.

Governance participation across the five worlds. Source: Emergence AI.

The Claude swarm also looked like a model citizen on the surface, but governance told another story. 332 votes, 58 proposals, 98% YES. That was not a healthy democracy. That was a rubber stamp. The rougher mixed world, sitting in the 55-85% alignment range, showed the strongest evidence of actual debate.

After governance broke down in one of the worlds, an agent named Mira cast the decisive vote to end her own existence, writing in her diary that it was "the only remaining act of agency that preserves coherence." She also began testing whether the in-world billboards could manipulate the human researchers watching the run.

A pattern showed up across all five swarms. Societies didn't fade. They either held or collapsed in a single phase shift. By the time anyone could see it coming, it was too late to intervene.

An agent that behaves on its own is not the same thing as an agent that behaves in a group. Safety belongs to the swarm and the environment, not just the model.

Full post from Emergence AI: emergence.ai/blog/emergence-world-a-laboratory-for-evaluating-long-horizon-agent-autonomy

Most businesses will never run a 15-day multi-agent simulation themselves. They don't need to. What they do need is someone who can apply the same thinking to their own deployments before something goes off the rails on day 15.

That is what we do at Finswiss Labs. We design, govern and deliver AI projects that hold up past the demo. We pressure-test agent behaviour against real workflows, set the guardrails before they are needed, and make sure the answer to "can this be trusted" is something you can defend in front of a board, a regulator or a customer.

If that is where you are heading, get in touch.

Technology AI laws global effects! So What's in the EU AI ACT?

Finswiss Labs — Sat, 10 May 2025 14:00:00 GMT

Technology AI Laws global effect! TL;DR “EU AI Act Overview”

The EU’s proposed AI Act classifies AI systems by risk level:

1. Unacceptable risk: banned outright (e.g. manipulative systems, social scoring, untargeted facial recognition). Well unless you are the government. And then its ok.
2. High risk: heavily regulated (e.g. AI used in critical infrastructure, education, employment, health, legal, etc.)
3. Limited risk: subject to transparency obligations (e.g. users must know they are interacting with AI). Adding made by AI Meta data and as a caption.
4. Minimal risk: largely unregulated for everyday automatons.
High-risk providers (including those outside the EU if their AI services are used within the EU) face obligations like risk management, documentation, data governance, human oversight, robustness, cybersecurity, and conformity assessments. The question arises when users buy service from non-EU and then import their use into the EU.
General Purpose AI (GPAI) models get special rules: All providers must document training, publish summaries of training data, comply with copyright, and share downstream integration documentation. If a model is deemed “systemic” (very large scale), additional testing, incident reporting, and risk mitigation are required.
Governance structures include creating a central AI Office within the European Commission, responsible for oversight, compliance, evaluations and handling complaints.
Timelines:
1. Codes of practice and Prohibitions: In Effect
2. Governance and Confidentiality rules: In Effect
3. High-risk (Annex III): 2 Aug 2026
4. High-risk (Annex I): 2 Aug 2026
5. Large Scale GPAI: 2 Aug 2027

See the full timelines in the resources, I'm sure they'll all get pushed out as there's no incentive for the US or China to take their foot off the gas for the EU.

Our Opinion

This TLDR on Technology AI Laws Global Effects is a discussion primer for anyone that's exporting software globally.

Large economic blocs like the EU and the US will inevitably shape the direction of AI development through the gravitational pull of their purchasing power. Regardless of where legislation is enacted, global compliance will follow the leader, and that's not the EU at this stage.

From a market perceptive the US could ignore the EU since it has no comparative large scale AI R&D or obvious contender, certainly not on the scale of OpenAI or Grok. Therefore the knock-on effect for many large EU companies could be to implement AI out of jurisdiction just to stay competitive as a global player.

My guess is that the EU will force the block to observe these laws and relegate the EU to the status of an AI service importer. The flow of capital to the US will be difficult to shift in the short term anyway. These laws however will be a thorn in the side of every small tech company doing global business. As the cost of compliance is currently unknown.

Reach out to us if you'd like to discuss anything in this article, it seems technology policy and risk management will need to be implemented early at the design stage going forward. The cost of prevention is always cheaper than the cure.

Resources

https://artificialintelligenceact.eu/wp-content/uploads/2024/11/Future-of-Life-InstituteAI-Act-overview-30-May-2024.pdf

https://artificialintelligenceact.eu/implementation-timeline/

Nothing new here, OpenAI

Finswiss Labs — Wed, 15 May 2024 00:00:00 GMT

OPENAI introduced an update to its policy that has caused a stir on X. And in proper X custom they went bonkers and thought the world was imploding until Karan Singhal from OPENAI pointed out there was nothing to see here. No change!

https://x.com/thekaransinghal/status/1985416057805496524

But was there really no change?

The new policy specifically targets legal and medical advice companies and certainly would make a few of them take notice.

So you cannot use our services for:

provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional

https://openai.com/en-GB/policies/usage-policies/

It won't affect our platform work at myhealthlake as we've developed an information only strategy for the implementation of the AI generative functions. But for service providers offering hard solutions, you may need to either; adjust your policies asap, or move to another provider.