China is likely to wake up to the AGI in May 2026 or to the fact that the US is weak in September 2025.

The Taiwan invasion is currently forecasted to start in 2027.  However, a comparable invasion done by Russia in late February 2022 was first forecasted as late as September 2021, implying that China may accelerate the invasion to just about six months after the decision.  

Consider the Scenario A when Taiwan is destroyed in November 2026 (18 months from May 2025) and Scenario B when this happens in March 2026 (10 months from May 2025). 

Compute estimates

The world's compute production is multiplied by 2.25x per year, and the AI-related share is multiplied by 1.5x per year.   

As of May^[1] 2025, OpenBrain's compute conducted 6.4E26 operations/month. DeepCent's compute conducted only 7.2E25 operations/month. These numbers are thought to rise to 9.8E26 and 1.2E26 by Sep 2026, to 1.5E27 and 2.1E26 by Jan 2026, to 2.2E27 and 3.6E26 by May 2026. Even without China's awakening or NVIDIA openly^[2] selling compute, DeepCent's compute is thought to increase five times a year. OpenBrain's compute is, as we know, forecasted to increase 3.4 times a year. If the USA and China were on different planets and this trend continued forever, then the two companies' compute would become equal in January 2031. 

Taiwan invasion forces both countries to use only domestic chips whose total production is about 10% or less of the former world market. In other words, the speed of receiving the compute by American AI companies is lowered from >90% to 5% or to 7.5%, i.e. 12–18  times. However, it may also prompt the US leaders to give OpenBrain the rivals' compute immediately (which increases OpenBrain's compute 2.5 times).

Therefore, if the time $t$ is the years that have passed since April or May 2025, OpenBrain's compute, expressed in ops/month, is $C_{OB}\left(t\right)=\{\begin{array}{cc}6.4\cdot {10}^{25}\cdot {1.5}^{3t}& \text{before the invasion}\\ 6.4\cdot {10}^{25}\cdot (\frac{{1.5}^{3t}}{N}+(1-1/N)\cdot {1.5}^{3t_{inv}})& \text{after the invasion}\end{array}$

$C_{OB}\left(t\right)=2.5\cdot 6.4\cdot {10}^{26}\cdot (\frac{{1.5}^{3t}}{N}+(1-1/N)\cdot {1.5}^{3t_{inv}})\text{after the unification and invasion}$

Next we suppose that DeepCent's compute increases exponentially without taking into account the fact that in May 2026 China decides to smuggle the chips from Taiwan or buy them wholesale, as NVIDIA advocates for some reason. After the awakening in May 2026 this optimistic for US scenario has $C_{DC}\left(t\right)=2.5\cdot 7.2\cdot {10}^{25}\cdot 5^t$

Progress estimates

The slowdown ending of the AI-2027 scenario has the creation of the robot economy begin in February or March 2028. By that point the AI systems provide a 150x speedup of the research. In the Race Ending similarly capable AI systems appear in November 2027, but a safety case appears only between December 2027 and June 2028. November 2027 is 30 months away from May 2025. But when do OpenBrain and DeepCent develop the systems of similar capabilities if the invasion occurs?

Anti-compute scenario

In order to estimate the impact of compute before the rise of superhuman coders and researchers, we quote the AI-2027 forecast:

About half of total progress historically has come from improved algorithms (which includes better ideas and new paradigms), the other half having come from scaled-up compute. So a 4x increase in the rate of algorithmic progress corresponds to a roughly 2x increase in the overall rate of progress.

Suppose that a company has compute $C\left(t\right)$ and has the AIs with capabilities $k\left(t\right)$. Then $k^{\prime }\left(t\right)=\sqrt{C\left(t\right)}f\left(k\right(t\left)\right)$, and  $\frac{k^{\prime }\left(t\right)}{f\left(k\right(t\left)\right)}=\sqrt{C\left(t\right)}$, meaning that the company's progress is defined by the integral ${\int }_{-\infty }^T\sqrt{C}dt$. 

# Let's extend the simulation to find when DeepCent eventually overtakes OpenBrain (if ever)
# for N = 18
from math import pow
# Constants
target_total_ops = 2e15
t_inv = 10  # invasion month (Mar 2026)
t_theft = 20 #theft month is Jan 2027
# Initial compute (in operations per month)
openbrain_initial = 6.4e26
deepcent_initial = 7.2e25
# Monthly growth factors
openbrain_growth = pow(1.5 ** 3, 1 / 12)  # (1.5)^3 per year
deepcent_growth = pow(5, 1 / 12)          # 5x per year
# Unknown parameter: N (slowdown factor post-invasion for OpenBrain)
# We'll leave it as a parameter and try several values
def compute_time_to_target(N):
  month = 0
  openbrain_compute = openbrain_initial
  deepcent_compute = deepcent_initial
  openbrain_produce = openbrain_initial*(openbrain_growth-1)
  deepcent_produce = deepcent_initial*(deepcent_growth-1)
  openbrain_ops = 0.0
  deepcent_ops = 0.0
  while (openbrain_ops < target_total_ops) and (deepcent_ops< target_total_ops):
      if month == 12:
          deepcent_compute *= 2.5
          deepcent_produce*=2.5 # DeepCent merges in May 2026
      if month == t_inv:
          openbrain_compute *= 2.5
          openbrain_produce*=2.5/N #OpenBrain merges right after the invasion
      if month == t_theft:
           deepcent_ops = max(openbrain_ops,deepcent_ops)
      openbrain_ops += pow(openbrain_compute, 1/2)
      deepcent_ops += pow(deepcent_compute, 1/2)
      openbrain_compute += openbrain_produce
      deepcent_compute += deepcent_produce
      openbrain_produce*=openbrain_growth
      deepcent_produce*=deepcent_growth
      month += 1
  print("OpenBrain did ")
  print(openbrain_ops)
  print(" DeepCent did ")
  print(deepcent_ops)
  return month
print(compute_time_to_target(18.0))

If Agent-2 wasn't stolen, then the anti-compute scenario would help the USA thrive despite the invasion by having the lead. 

1. However, if Agent-2 is stolen and Taiwan is invaded in March 2026, then the period of American leadership will shrink to three months, and the slowdown ending is impossible or highly unlikely. If Taiwan is invaded in November 2026 and Agent-2 is stolen, then the leadership is eight months; however, the slowdown might cause DeepCent to outrace OpenBrain and to end up with a misaligned AI released a bit earlier than the US.
2. The scenario is also vulnerable to the fact that superhuman AIs require more compute and/or high-quality data (which is likely to be AI-created) to train.
3. If superhuman AI researchers arrive, then the amount of intellectual labor these researchers invest is proportional to the compute spent on them. This moves us to the next scenario.

Pro-compute scenario

# Let's extend the simulation to find when DeepCent eventually overtakes OpenBrain (if ever)
# for N = 18
from math import pow
# Constants
target_total_ops = 1e29
t_inv = 18  # invasion month (Nov 2026, 1.5 years after May 2025)
t_theft = 20 #theft month
# Initial compute (in operations per month)
openbrain_initial = 6.4e26
deepcent_initial = 7.2e25
# Monthly growth factors
openbrain_growth = pow(1.5 ** 3, 1 / 12)  # (1.5)^3 per year
deepcent_growth = pow(5, 1 / 12)          # 5x per year
# Unknown parameter: N (slowdown factor post-invasion for OpenBrain)
# We'll leave it as a parameter and try several values
def compute_time_to_target(N):
  month = 0
  openbrain_compute = openbrain_initial
  deepcent_compute = deepcent_initial
  openbrain_produce = openbrain_initial*(openbrain_growth-1)
  deepcent_produce = deepcent_initial*(deepcent_growth-1)
  openbrain_ops = 0.0
  deepcent_ops = 0.0
  while (openbrain_ops < target_total_ops) and (deepcent_ops< target_total_ops):
      if month == 12:
          deepcent_compute *= 2.5
          deepcent_produce*=2.5 # DeepCent merges in May 2026
      if month == t_inv:
          openbrain_compute *= 2.5
          openbrain_produce*=2.5/N
      if month == t_theft:
           deepcent_ops = max(openbrain_ops,deepcent_ops)
      openbrain_ops += pow(openbrain_compute, 1)
      deepcent_ops += pow(deepcent_compute, 1)
      openbrain_compute += openbrain_produce
      deepcent_compute += deepcent_produce
      openbrain_produce*=openbrain_growth
      deepcent_produce*=deepcent_growth
      month += 1
  print("OpenBrain did ")
  print(openbrain_ops)
  print(" DeepCent did ")
  print(deepcent_ops)
  return month
print(compute_time_to_target(18.0))

The simulation ensures that if the invasion is done in November 2026, then OpenBrain and DeepCent do 3E29 operations in 41 months since May 2025, or in December 2028. While this is too late, this estimate relies on OpenBrain merging. If OpenBrain fails to merge, then they do 1E29 operations in 33 months, as does DeepCent. Any problems with the US economy threaten to accelerate DeepCent or slow OpenBrain to the point of decreasing the chances for the slowdown.

If the invasion happens in March 2026 (10 months from May 2025), then merging causes OpenBrain and DeepCent to do the same amount of compute in 37 months, and the slowdown ending becomes outright impossible.

Conclusion

Therefore, the invasion is likely to cause any problems with the US economy that could slow OpenBrain down to become fatal for American leadership or for the possibility of the slowdown ending where a powerful aligned AI constrains a weak misaligned one. The only scenarios where mankind survives are the Race Ending where the Agents are actually aligned^[3] in a different way, a mutual slowdown (e.g. if the rogue AIs are hard to deal with), or the idea that one side decides to slow down and the weak aligned AGI^[4] constrains the powerful misaligned one. 

1. ^{^}

   In June the AI-2027 forecast was updated to have OpenBrain's compute equal to 9.8E25 ops/month in August 2025, not in September 2025. If I remember correctly, the forecast had the same numbers in April and June, but was shifted one month later in May.

2. ^{^}

   The news about NVIDIA's intentions reached the public in May, after the AI-2027 forecast was released.

3. ^{^}

   For example, if their goal is to prevent mankind from being destroyed and from becoming parasites. In this case, instead of killing all humans, the AIs leave them without the potential to use the AIs in obviously parasitic goals, which also puts off the possibility of the Deep Utopia or the Intelligence Curse until humans or superbabies solve everything they can solve by themselves. 

4. ^{^}

   Similar scenarios are proposed in the posts by Arkhos Winter and me. However, I envision the chance that American culture or the American race against inner troubles causes the American AI to be misaligned and more powerful than the Chinese one who is aligned, but weaker. While the two AIs are likely to achieve a compromiss, like in the AI-2027 scenario, the misaligned AI gets FAR more resources.